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/RuntimeLibcalls.h" 32 #include "llvm/CodeGen/StackMaps.h" 33 #include "llvm/CodeGen/TargetLowering.h" 34 #include "llvm/CodeGen/TargetOpcodes.h" 35 #include "llvm/CodeGen/TargetRegisterInfo.h" 36 #include "llvm/CodeGen/ValueTypes.h" 37 #include "llvm/IR/Attributes.h" 38 #include "llvm/IR/CallingConv.h" 39 #include "llvm/IR/DataLayout.h" 40 #include "llvm/IR/DerivedTypes.h" 41 #include "llvm/IR/Function.h" 42 #include "llvm/IR/GlobalValue.h" 43 #include "llvm/IR/GlobalVariable.h" 44 #include "llvm/IR/IRBuilder.h" 45 #include "llvm/IR/Module.h" 46 #include "llvm/IR/Type.h" 47 #include "llvm/Support/BranchProbability.h" 48 #include "llvm/Support/Casting.h" 49 #include "llvm/Support/CommandLine.h" 50 #include "llvm/Support/Compiler.h" 51 #include "llvm/Support/ErrorHandling.h" 52 #include "llvm/Support/MachineValueType.h" 53 #include "llvm/Support/MathExtras.h" 54 #include "llvm/Target/TargetMachine.h" 55 #include <algorithm> 56 #include <cassert> 57 #include <cstddef> 58 #include <cstdint> 59 #include <cstring> 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 static bool darwinHasSinCos(const Triple &TT) { 93 assert(TT.isOSDarwin() && "should be called with darwin triple"); 94 // Don't bother with 32 bit x86. 95 if (TT.getArch() == Triple::x86) 96 return false; 97 // Macos < 10.9 has no sincos_stret. 98 if (TT.isMacOSX()) 99 return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit(); 100 // iOS < 7.0 has no sincos_stret. 101 if (TT.isiOS()) 102 return !TT.isOSVersionLT(7, 0); 103 // Any other darwin such as WatchOS/TvOS is new enough. 104 return true; 105 } 106 107 // Although this default value is arbitrary, it is not random. It is assumed 108 // that a condition that evaluates the same way by a higher percentage than this 109 // is best represented as control flow. Therefore, the default value N should be 110 // set such that the win from N% correct executions is greater than the loss 111 // from (100 - N)% mispredicted executions for the majority of intended targets. 112 static cl::opt<int> MinPercentageForPredictableBranch( 113 "min-predictable-branch", cl::init(99), 114 cl::desc("Minimum percentage (0-100) that a condition must be either true " 115 "or false to assume that the condition is predictable"), 116 cl::Hidden); 117 118 void TargetLoweringBase::InitLibcalls(const Triple &TT) { 119 #define HANDLE_LIBCALL(code, name) \ 120 setLibcallName(RTLIB::code, name); 121 #include "llvm/CodeGen/RuntimeLibcalls.def" 122 #undef HANDLE_LIBCALL 123 // Initialize calling conventions to their default. 124 for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC) 125 setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C); 126 127 // A few names are different on particular architectures or environments. 128 if (TT.isOSDarwin()) { 129 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead 130 // of the gnueabi-style __gnu_*_ieee. 131 // FIXME: What about other targets? 132 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2"); 133 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2"); 134 135 // Some darwins have an optimized __bzero/bzero function. 136 switch (TT.getArch()) { 137 case Triple::x86: 138 case Triple::x86_64: 139 if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6)) 140 setLibcallName(RTLIB::BZERO, "__bzero"); 141 break; 142 case Triple::aarch64: 143 setLibcallName(RTLIB::BZERO, "bzero"); 144 break; 145 default: 146 break; 147 } 148 149 if (darwinHasSinCos(TT)) { 150 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret"); 151 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret"); 152 if (TT.isWatchABI()) { 153 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32, 154 CallingConv::ARM_AAPCS_VFP); 155 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64, 156 CallingConv::ARM_AAPCS_VFP); 157 } 158 } 159 } else { 160 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee"); 161 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee"); 162 } 163 164 if (TT.isGNUEnvironment() || TT.isOSFuchsia()) { 165 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 166 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 167 setLibcallName(RTLIB::SINCOS_F80, "sincosl"); 168 setLibcallName(RTLIB::SINCOS_F128, "sincosl"); 169 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl"); 170 } 171 172 if (TT.isOSOpenBSD()) { 173 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr); 174 } 175 } 176 177 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 178 /// UNKNOWN_LIBCALL if there is none. 179 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 180 if (OpVT == MVT::f16) { 181 if (RetVT == MVT::f32) 182 return FPEXT_F16_F32; 183 } else if (OpVT == MVT::f32) { 184 if (RetVT == MVT::f64) 185 return FPEXT_F32_F64; 186 if (RetVT == MVT::f128) 187 return FPEXT_F32_F128; 188 if (RetVT == MVT::ppcf128) 189 return FPEXT_F32_PPCF128; 190 } else if (OpVT == MVT::f64) { 191 if (RetVT == MVT::f128) 192 return FPEXT_F64_F128; 193 else if (RetVT == MVT::ppcf128) 194 return FPEXT_F64_PPCF128; 195 } else if (OpVT == MVT::f80) { 196 if (RetVT == MVT::f128) 197 return FPEXT_F80_F128; 198 } 199 200 return UNKNOWN_LIBCALL; 201 } 202 203 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 204 /// UNKNOWN_LIBCALL if there is none. 205 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 206 if (RetVT == MVT::f16) { 207 if (OpVT == MVT::f32) 208 return FPROUND_F32_F16; 209 if (OpVT == MVT::f64) 210 return FPROUND_F64_F16; 211 if (OpVT == MVT::f80) 212 return FPROUND_F80_F16; 213 if (OpVT == MVT::f128) 214 return FPROUND_F128_F16; 215 if (OpVT == MVT::ppcf128) 216 return FPROUND_PPCF128_F16; 217 } else if (RetVT == MVT::f32) { 218 if (OpVT == MVT::f64) 219 return FPROUND_F64_F32; 220 if (OpVT == MVT::f80) 221 return FPROUND_F80_F32; 222 if (OpVT == MVT::f128) 223 return FPROUND_F128_F32; 224 if (OpVT == MVT::ppcf128) 225 return FPROUND_PPCF128_F32; 226 } else if (RetVT == MVT::f64) { 227 if (OpVT == MVT::f80) 228 return FPROUND_F80_F64; 229 if (OpVT == MVT::f128) 230 return FPROUND_F128_F64; 231 if (OpVT == MVT::ppcf128) 232 return FPROUND_PPCF128_F64; 233 } else if (RetVT == MVT::f80) { 234 if (OpVT == MVT::f128) 235 return FPROUND_F128_F80; 236 } 237 238 return UNKNOWN_LIBCALL; 239 } 240 241 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 242 /// UNKNOWN_LIBCALL if there is none. 243 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 244 if (OpVT == MVT::f32) { 245 if (RetVT == MVT::i32) 246 return FPTOSINT_F32_I32; 247 if (RetVT == MVT::i64) 248 return FPTOSINT_F32_I64; 249 if (RetVT == MVT::i128) 250 return FPTOSINT_F32_I128; 251 } else if (OpVT == MVT::f64) { 252 if (RetVT == MVT::i32) 253 return FPTOSINT_F64_I32; 254 if (RetVT == MVT::i64) 255 return FPTOSINT_F64_I64; 256 if (RetVT == MVT::i128) 257 return FPTOSINT_F64_I128; 258 } else if (OpVT == MVT::f80) { 259 if (RetVT == MVT::i32) 260 return FPTOSINT_F80_I32; 261 if (RetVT == MVT::i64) 262 return FPTOSINT_F80_I64; 263 if (RetVT == MVT::i128) 264 return FPTOSINT_F80_I128; 265 } else if (OpVT == MVT::f128) { 266 if (RetVT == MVT::i32) 267 return FPTOSINT_F128_I32; 268 if (RetVT == MVT::i64) 269 return FPTOSINT_F128_I64; 270 if (RetVT == MVT::i128) 271 return FPTOSINT_F128_I128; 272 } else if (OpVT == MVT::ppcf128) { 273 if (RetVT == MVT::i32) 274 return FPTOSINT_PPCF128_I32; 275 if (RetVT == MVT::i64) 276 return FPTOSINT_PPCF128_I64; 277 if (RetVT == MVT::i128) 278 return FPTOSINT_PPCF128_I128; 279 } 280 return UNKNOWN_LIBCALL; 281 } 282 283 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 284 /// UNKNOWN_LIBCALL if there is none. 285 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 286 if (OpVT == MVT::f32) { 287 if (RetVT == MVT::i32) 288 return FPTOUINT_F32_I32; 289 if (RetVT == MVT::i64) 290 return FPTOUINT_F32_I64; 291 if (RetVT == MVT::i128) 292 return FPTOUINT_F32_I128; 293 } else if (OpVT == MVT::f64) { 294 if (RetVT == MVT::i32) 295 return FPTOUINT_F64_I32; 296 if (RetVT == MVT::i64) 297 return FPTOUINT_F64_I64; 298 if (RetVT == MVT::i128) 299 return FPTOUINT_F64_I128; 300 } else if (OpVT == MVT::f80) { 301 if (RetVT == MVT::i32) 302 return FPTOUINT_F80_I32; 303 if (RetVT == MVT::i64) 304 return FPTOUINT_F80_I64; 305 if (RetVT == MVT::i128) 306 return FPTOUINT_F80_I128; 307 } else if (OpVT == MVT::f128) { 308 if (RetVT == MVT::i32) 309 return FPTOUINT_F128_I32; 310 if (RetVT == MVT::i64) 311 return FPTOUINT_F128_I64; 312 if (RetVT == MVT::i128) 313 return FPTOUINT_F128_I128; 314 } else if (OpVT == MVT::ppcf128) { 315 if (RetVT == MVT::i32) 316 return FPTOUINT_PPCF128_I32; 317 if (RetVT == MVT::i64) 318 return FPTOUINT_PPCF128_I64; 319 if (RetVT == MVT::i128) 320 return FPTOUINT_PPCF128_I128; 321 } 322 return UNKNOWN_LIBCALL; 323 } 324 325 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 326 /// UNKNOWN_LIBCALL if there is none. 327 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 328 if (OpVT == MVT::i32) { 329 if (RetVT == MVT::f32) 330 return SINTTOFP_I32_F32; 331 if (RetVT == MVT::f64) 332 return SINTTOFP_I32_F64; 333 if (RetVT == MVT::f80) 334 return SINTTOFP_I32_F80; 335 if (RetVT == MVT::f128) 336 return SINTTOFP_I32_F128; 337 if (RetVT == MVT::ppcf128) 338 return SINTTOFP_I32_PPCF128; 339 } else if (OpVT == MVT::i64) { 340 if (RetVT == MVT::f32) 341 return SINTTOFP_I64_F32; 342 if (RetVT == MVT::f64) 343 return SINTTOFP_I64_F64; 344 if (RetVT == MVT::f80) 345 return SINTTOFP_I64_F80; 346 if (RetVT == MVT::f128) 347 return SINTTOFP_I64_F128; 348 if (RetVT == MVT::ppcf128) 349 return SINTTOFP_I64_PPCF128; 350 } else if (OpVT == MVT::i128) { 351 if (RetVT == MVT::f32) 352 return SINTTOFP_I128_F32; 353 if (RetVT == MVT::f64) 354 return SINTTOFP_I128_F64; 355 if (RetVT == MVT::f80) 356 return SINTTOFP_I128_F80; 357 if (RetVT == MVT::f128) 358 return SINTTOFP_I128_F128; 359 if (RetVT == MVT::ppcf128) 360 return SINTTOFP_I128_PPCF128; 361 } 362 return UNKNOWN_LIBCALL; 363 } 364 365 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 366 /// UNKNOWN_LIBCALL if there is none. 367 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 368 if (OpVT == MVT::i32) { 369 if (RetVT == MVT::f32) 370 return UINTTOFP_I32_F32; 371 if (RetVT == MVT::f64) 372 return UINTTOFP_I32_F64; 373 if (RetVT == MVT::f80) 374 return UINTTOFP_I32_F80; 375 if (RetVT == MVT::f128) 376 return UINTTOFP_I32_F128; 377 if (RetVT == MVT::ppcf128) 378 return UINTTOFP_I32_PPCF128; 379 } else if (OpVT == MVT::i64) { 380 if (RetVT == MVT::f32) 381 return UINTTOFP_I64_F32; 382 if (RetVT == MVT::f64) 383 return UINTTOFP_I64_F64; 384 if (RetVT == MVT::f80) 385 return UINTTOFP_I64_F80; 386 if (RetVT == MVT::f128) 387 return UINTTOFP_I64_F128; 388 if (RetVT == MVT::ppcf128) 389 return UINTTOFP_I64_PPCF128; 390 } else if (OpVT == MVT::i128) { 391 if (RetVT == MVT::f32) 392 return UINTTOFP_I128_F32; 393 if (RetVT == MVT::f64) 394 return UINTTOFP_I128_F64; 395 if (RetVT == MVT::f80) 396 return UINTTOFP_I128_F80; 397 if (RetVT == MVT::f128) 398 return UINTTOFP_I128_F128; 399 if (RetVT == MVT::ppcf128) 400 return UINTTOFP_I128_PPCF128; 401 } 402 return UNKNOWN_LIBCALL; 403 } 404 405 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 406 #define OP_TO_LIBCALL(Name, Enum) \ 407 case Name: \ 408 switch (VT.SimpleTy) { \ 409 default: \ 410 return UNKNOWN_LIBCALL; \ 411 case MVT::i8: \ 412 return Enum##_1; \ 413 case MVT::i16: \ 414 return Enum##_2; \ 415 case MVT::i32: \ 416 return Enum##_4; \ 417 case MVT::i64: \ 418 return Enum##_8; \ 419 case MVT::i128: \ 420 return Enum##_16; \ 421 } 422 423 switch (Opc) { 424 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 425 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 426 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 427 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 428 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 429 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 430 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 431 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 432 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 433 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 434 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 435 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 436 } 437 438 #undef OP_TO_LIBCALL 439 440 return UNKNOWN_LIBCALL; 441 } 442 443 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 444 switch (ElementSize) { 445 case 1: 446 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 447 case 2: 448 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 449 case 4: 450 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 451 case 8: 452 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 453 case 16: 454 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 455 default: 456 return UNKNOWN_LIBCALL; 457 } 458 } 459 460 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 461 switch (ElementSize) { 462 case 1: 463 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 464 case 2: 465 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 466 case 4: 467 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 468 case 8: 469 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 470 case 16: 471 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 472 default: 473 return UNKNOWN_LIBCALL; 474 } 475 } 476 477 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 478 switch (ElementSize) { 479 case 1: 480 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 481 case 2: 482 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 483 case 4: 484 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 485 case 8: 486 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 487 case 16: 488 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 489 default: 490 return UNKNOWN_LIBCALL; 491 } 492 } 493 494 /// InitCmpLibcallCCs - Set default comparison libcall CC. 495 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 496 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 497 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 498 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 499 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 500 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 501 CCs[RTLIB::UNE_F32] = ISD::SETNE; 502 CCs[RTLIB::UNE_F64] = ISD::SETNE; 503 CCs[RTLIB::UNE_F128] = ISD::SETNE; 504 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 505 CCs[RTLIB::OGE_F32] = ISD::SETGE; 506 CCs[RTLIB::OGE_F64] = ISD::SETGE; 507 CCs[RTLIB::OGE_F128] = ISD::SETGE; 508 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 509 CCs[RTLIB::OLT_F32] = ISD::SETLT; 510 CCs[RTLIB::OLT_F64] = ISD::SETLT; 511 CCs[RTLIB::OLT_F128] = ISD::SETLT; 512 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 513 CCs[RTLIB::OLE_F32] = ISD::SETLE; 514 CCs[RTLIB::OLE_F64] = ISD::SETLE; 515 CCs[RTLIB::OLE_F128] = ISD::SETLE; 516 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 517 CCs[RTLIB::OGT_F32] = ISD::SETGT; 518 CCs[RTLIB::OGT_F64] = ISD::SETGT; 519 CCs[RTLIB::OGT_F128] = ISD::SETGT; 520 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 521 CCs[RTLIB::UO_F32] = ISD::SETNE; 522 CCs[RTLIB::UO_F64] = ISD::SETNE; 523 CCs[RTLIB::UO_F128] = ISD::SETNE; 524 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 525 CCs[RTLIB::O_F32] = ISD::SETEQ; 526 CCs[RTLIB::O_F64] = ISD::SETEQ; 527 CCs[RTLIB::O_F128] = ISD::SETEQ; 528 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 529 } 530 531 /// NOTE: The TargetMachine owns TLOF. 532 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 533 initActions(); 534 535 // Perform these initializations only once. 536 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 537 MaxLoadsPerMemcmp = 8; 538 MaxGluedStoresPerMemcpy = 0; 539 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 540 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 541 UseUnderscoreSetJmp = false; 542 UseUnderscoreLongJmp = false; 543 HasMultipleConditionRegisters = false; 544 HasExtractBitsInsn = false; 545 JumpIsExpensive = JumpIsExpensiveOverride; 546 PredictableSelectIsExpensive = false; 547 EnableExtLdPromotion = false; 548 HasFloatingPointExceptions = true; 549 StackPointerRegisterToSaveRestore = 0; 550 BooleanContents = UndefinedBooleanContent; 551 BooleanFloatContents = UndefinedBooleanContent; 552 BooleanVectorContents = UndefinedBooleanContent; 553 SchedPreferenceInfo = Sched::ILP; 554 JumpBufSize = 0; 555 JumpBufAlignment = 0; 556 MinFunctionAlignment = 0; 557 PrefFunctionAlignment = 0; 558 PrefLoopAlignment = 0; 559 GatherAllAliasesMaxDepth = 18; 560 MinStackArgumentAlignment = 1; 561 // TODO: the default will be switched to 0 in the next commit, along 562 // with the Target-specific changes necessary. 563 MaxAtomicSizeInBitsSupported = 1024; 564 565 MinCmpXchgSizeInBits = 0; 566 SupportsUnalignedAtomics = false; 567 568 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 569 570 InitLibcalls(TM.getTargetTriple()); 571 InitCmpLibcallCCs(CmpLibcallCCs); 572 } 573 574 void TargetLoweringBase::initActions() { 575 // All operations default to being supported. 576 memset(OpActions, 0, sizeof(OpActions)); 577 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 578 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 579 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 580 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 581 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 582 std::fill(std::begin(TargetDAGCombineArray), 583 std::end(TargetDAGCombineArray), 0); 584 585 // Set default actions for various operations. 586 for (MVT VT : MVT::all_valuetypes()) { 587 // Default all indexed load / store to expand. 588 for (unsigned IM = (unsigned)ISD::PRE_INC; 589 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 590 setIndexedLoadAction(IM, VT, Expand); 591 setIndexedStoreAction(IM, VT, Expand); 592 } 593 594 // Most backends expect to see the node which just returns the value loaded. 595 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 596 597 // These operations default to expand. 598 setOperationAction(ISD::FGETSIGN, VT, Expand); 599 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 600 setOperationAction(ISD::FMINNUM, VT, Expand); 601 setOperationAction(ISD::FMAXNUM, VT, Expand); 602 setOperationAction(ISD::FMINNAN, VT, Expand); 603 setOperationAction(ISD::FMAXNAN, VT, Expand); 604 setOperationAction(ISD::FMAD, VT, Expand); 605 setOperationAction(ISD::SMIN, VT, Expand); 606 setOperationAction(ISD::SMAX, VT, Expand); 607 setOperationAction(ISD::UMIN, VT, Expand); 608 setOperationAction(ISD::UMAX, VT, Expand); 609 setOperationAction(ISD::ABS, VT, Expand); 610 611 // Overflow operations default to expand 612 setOperationAction(ISD::SADDO, VT, Expand); 613 setOperationAction(ISD::SSUBO, VT, Expand); 614 setOperationAction(ISD::UADDO, VT, Expand); 615 setOperationAction(ISD::USUBO, VT, Expand); 616 setOperationAction(ISD::SMULO, VT, Expand); 617 setOperationAction(ISD::UMULO, VT, Expand); 618 619 // ADDCARRY operations default to expand 620 setOperationAction(ISD::ADDCARRY, VT, Expand); 621 setOperationAction(ISD::SUBCARRY, VT, Expand); 622 setOperationAction(ISD::SETCCCARRY, VT, Expand); 623 624 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 625 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 626 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 627 628 setOperationAction(ISD::BITREVERSE, VT, Expand); 629 630 // These library functions default to expand. 631 setOperationAction(ISD::FROUND, VT, Expand); 632 setOperationAction(ISD::FPOWI, VT, Expand); 633 634 // These operations default to expand for vector types. 635 if (VT.isVector()) { 636 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 637 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 638 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 639 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 640 } 641 642 // For most targets @llvm.get.dynamic.area.offset just returns 0. 643 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 644 } 645 646 // Most targets ignore the @llvm.prefetch intrinsic. 647 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 648 649 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 650 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 651 652 // ConstantFP nodes default to expand. Targets can either change this to 653 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 654 // to optimize expansions for certain constants. 655 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 656 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 657 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 658 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 659 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 660 661 // These library functions default to expand. 662 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 663 setOperationAction(ISD::FLOG , VT, Expand); 664 setOperationAction(ISD::FLOG2, VT, Expand); 665 setOperationAction(ISD::FLOG10, VT, Expand); 666 setOperationAction(ISD::FEXP , VT, Expand); 667 setOperationAction(ISD::FEXP2, VT, Expand); 668 setOperationAction(ISD::FFLOOR, VT, Expand); 669 setOperationAction(ISD::FNEARBYINT, VT, Expand); 670 setOperationAction(ISD::FCEIL, VT, Expand); 671 setOperationAction(ISD::FRINT, VT, Expand); 672 setOperationAction(ISD::FTRUNC, VT, Expand); 673 setOperationAction(ISD::FROUND, VT, Expand); 674 } 675 676 // Default ISD::TRAP to expand (which turns it into abort). 677 setOperationAction(ISD::TRAP, MVT::Other, Expand); 678 679 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 680 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 681 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 682 } 683 684 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 685 EVT) const { 686 return MVT::getIntegerVT(8 * DL.getPointerSize(0)); 687 } 688 689 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 690 bool LegalTypes) const { 691 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 692 if (LHSTy.isVector()) 693 return LHSTy; 694 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 695 : getPointerTy(DL); 696 } 697 698 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 699 assert(isTypeLegal(VT)); 700 switch (Op) { 701 default: 702 return false; 703 case ISD::SDIV: 704 case ISD::UDIV: 705 case ISD::SREM: 706 case ISD::UREM: 707 return true; 708 } 709 } 710 711 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 712 // If the command-line option was specified, ignore this request. 713 if (!JumpIsExpensiveOverride.getNumOccurrences()) 714 JumpIsExpensive = isExpensive; 715 } 716 717 TargetLoweringBase::LegalizeKind 718 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 719 // If this is a simple type, use the ComputeRegisterProp mechanism. 720 if (VT.isSimple()) { 721 MVT SVT = VT.getSimpleVT(); 722 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 723 MVT NVT = TransformToType[SVT.SimpleTy]; 724 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 725 726 assert((LA == TypeLegal || LA == TypeSoftenFloat || 727 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) && 728 "Promote may not follow Expand or Promote"); 729 730 if (LA == TypeSplitVector) 731 return LegalizeKind(LA, 732 EVT::getVectorVT(Context, SVT.getVectorElementType(), 733 SVT.getVectorNumElements() / 2)); 734 if (LA == TypeScalarizeVector) 735 return LegalizeKind(LA, SVT.getVectorElementType()); 736 return LegalizeKind(LA, NVT); 737 } 738 739 // Handle Extended Scalar Types. 740 if (!VT.isVector()) { 741 assert(VT.isInteger() && "Float types must be simple"); 742 unsigned BitSize = VT.getSizeInBits(); 743 // First promote to a power-of-two size, then expand if necessary. 744 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 745 EVT NVT = VT.getRoundIntegerType(Context); 746 assert(NVT != VT && "Unable to round integer VT"); 747 LegalizeKind NextStep = getTypeConversion(Context, NVT); 748 // Avoid multi-step promotion. 749 if (NextStep.first == TypePromoteInteger) 750 return NextStep; 751 // Return rounded integer type. 752 return LegalizeKind(TypePromoteInteger, NVT); 753 } 754 755 return LegalizeKind(TypeExpandInteger, 756 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 757 } 758 759 // Handle vector types. 760 unsigned NumElts = VT.getVectorNumElements(); 761 EVT EltVT = VT.getVectorElementType(); 762 763 // Vectors with only one element are always scalarized. 764 if (NumElts == 1) 765 return LegalizeKind(TypeScalarizeVector, EltVT); 766 767 // Try to widen vector elements until the element type is a power of two and 768 // promote it to a legal type later on, for example: 769 // <3 x i8> -> <4 x i8> -> <4 x i32> 770 if (EltVT.isInteger()) { 771 // Vectors with a number of elements that is not a power of two are always 772 // widened, for example <3 x i8> -> <4 x i8>. 773 if (!VT.isPow2VectorType()) { 774 NumElts = (unsigned)NextPowerOf2(NumElts); 775 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 776 return LegalizeKind(TypeWidenVector, NVT); 777 } 778 779 // Examine the element type. 780 LegalizeKind LK = getTypeConversion(Context, EltVT); 781 782 // If type is to be expanded, split the vector. 783 // <4 x i140> -> <2 x i140> 784 if (LK.first == TypeExpandInteger) 785 return LegalizeKind(TypeSplitVector, 786 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 787 788 // Promote the integer element types until a legal vector type is found 789 // or until the element integer type is too big. If a legal type was not 790 // found, fallback to the usual mechanism of widening/splitting the 791 // vector. 792 EVT OldEltVT = EltVT; 793 while (true) { 794 // Increase the bitwidth of the element to the next pow-of-two 795 // (which is greater than 8 bits). 796 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 797 .getRoundIntegerType(Context); 798 799 // Stop trying when getting a non-simple element type. 800 // Note that vector elements may be greater than legal vector element 801 // types. Example: X86 XMM registers hold 64bit element on 32bit 802 // systems. 803 if (!EltVT.isSimple()) 804 break; 805 806 // Build a new vector type and check if it is legal. 807 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 808 // Found a legal promoted vector type. 809 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 810 return LegalizeKind(TypePromoteInteger, 811 EVT::getVectorVT(Context, EltVT, NumElts)); 812 } 813 814 // Reset the type to the unexpanded type if we did not find a legal vector 815 // type with a promoted vector element type. 816 EltVT = OldEltVT; 817 } 818 819 // Try to widen the vector until a legal type is found. 820 // If there is no wider legal type, split the vector. 821 while (true) { 822 // Round up to the next power of 2. 823 NumElts = (unsigned)NextPowerOf2(NumElts); 824 825 // If there is no simple vector type with this many elements then there 826 // cannot be a larger legal vector type. Note that this assumes that 827 // there are no skipped intermediate vector types in the simple types. 828 if (!EltVT.isSimple()) 829 break; 830 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 831 if (LargerVector == MVT()) 832 break; 833 834 // If this type is legal then widen the vector. 835 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 836 return LegalizeKind(TypeWidenVector, LargerVector); 837 } 838 839 // Widen odd vectors to next power of two. 840 if (!VT.isPow2VectorType()) { 841 EVT NVT = VT.getPow2VectorType(Context); 842 return LegalizeKind(TypeWidenVector, NVT); 843 } 844 845 // Vectors with illegal element types are expanded. 846 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 847 return LegalizeKind(TypeSplitVector, NVT); 848 } 849 850 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 851 unsigned &NumIntermediates, 852 MVT &RegisterVT, 853 TargetLoweringBase *TLI) { 854 // Figure out the right, legal destination reg to copy into. 855 unsigned NumElts = VT.getVectorNumElements(); 856 MVT EltTy = VT.getVectorElementType(); 857 858 unsigned NumVectorRegs = 1; 859 860 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 861 // could break down into LHS/RHS like LegalizeDAG does. 862 if (!isPowerOf2_32(NumElts)) { 863 NumVectorRegs = NumElts; 864 NumElts = 1; 865 } 866 867 // Divide the input until we get to a supported size. This will always 868 // end with a scalar if the target doesn't support vectors. 869 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 870 NumElts >>= 1; 871 NumVectorRegs <<= 1; 872 } 873 874 NumIntermediates = NumVectorRegs; 875 876 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 877 if (!TLI->isTypeLegal(NewVT)) 878 NewVT = EltTy; 879 IntermediateVT = NewVT; 880 881 unsigned NewVTSize = NewVT.getSizeInBits(); 882 883 // Convert sizes such as i33 to i64. 884 if (!isPowerOf2_32(NewVTSize)) 885 NewVTSize = NextPowerOf2(NewVTSize); 886 887 MVT DestVT = TLI->getRegisterType(NewVT); 888 RegisterVT = DestVT; 889 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 890 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 891 892 // Otherwise, promotion or legal types use the same number of registers as 893 // the vector decimated to the appropriate level. 894 return NumVectorRegs; 895 } 896 897 /// isLegalRC - Return true if the value types that can be represented by the 898 /// specified register class are all legal. 899 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 900 const TargetRegisterClass &RC) const { 901 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 902 if (isTypeLegal(*I)) 903 return true; 904 return false; 905 } 906 907 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 908 /// sequence of memory operands that is recognized by PrologEpilogInserter. 909 MachineBasicBlock * 910 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 911 MachineBasicBlock *MBB) const { 912 MachineInstr *MI = &InitialMI; 913 MachineFunction &MF = *MI->getMF(); 914 MachineFrameInfo &MFI = MF.getFrameInfo(); 915 916 // We're handling multiple types of operands here: 917 // PATCHPOINT MetaArgs - live-in, read only, direct 918 // STATEPOINT Deopt Spill - live-through, read only, indirect 919 // STATEPOINT Deopt Alloca - live-through, read only, direct 920 // (We're currently conservative and mark the deopt slots read/write in 921 // practice.) 922 // STATEPOINT GC Spill - live-through, read/write, indirect 923 // STATEPOINT GC Alloca - live-through, read/write, direct 924 // The live-in vs live-through is handled already (the live through ones are 925 // all stack slots), but we need to handle the different type of stackmap 926 // operands and memory effects here. 927 928 // MI changes inside this loop as we grow operands. 929 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 930 MachineOperand &MO = MI->getOperand(OperIdx); 931 if (!MO.isFI()) 932 continue; 933 934 // foldMemoryOperand builds a new MI after replacing a single FI operand 935 // with the canonical set of five x86 addressing-mode operands. 936 int FI = MO.getIndex(); 937 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 938 939 // Copy operands before the frame-index. 940 for (unsigned i = 0; i < OperIdx; ++i) 941 MIB.add(MI->getOperand(i)); 942 // Add frame index operands recognized by stackmaps.cpp 943 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 944 // indirect-mem-ref tag, size, #FI, offset. 945 // Used for spills inserted by StatepointLowering. This codepath is not 946 // used for patchpoints/stackmaps at all, for these spilling is done via 947 // foldMemoryOperand callback only. 948 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 949 MIB.addImm(StackMaps::IndirectMemRefOp); 950 MIB.addImm(MFI.getObjectSize(FI)); 951 MIB.add(MI->getOperand(OperIdx)); 952 MIB.addImm(0); 953 } else { 954 // direct-mem-ref tag, #FI, offset. 955 // Used by patchpoint, and direct alloca arguments to statepoints 956 MIB.addImm(StackMaps::DirectMemRefOp); 957 MIB.add(MI->getOperand(OperIdx)); 958 MIB.addImm(0); 959 } 960 // Copy the operands after the frame index. 961 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 962 MIB.add(MI->getOperand(i)); 963 964 // Inherit previous memory operands. 965 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 966 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 967 968 // Add a new memory operand for this FI. 969 assert(MFI.getObjectOffset(FI) != -1); 970 971 auto Flags = MachineMemOperand::MOLoad; 972 if (MI->getOpcode() == TargetOpcode::STATEPOINT) { 973 Flags |= MachineMemOperand::MOStore; 974 Flags |= MachineMemOperand::MOVolatile; 975 } 976 MachineMemOperand *MMO = MF.getMachineMemOperand( 977 MachinePointerInfo::getFixedStack(MF, FI), Flags, 978 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 979 MIB->addMemOperand(MF, MMO); 980 981 // Replace the instruction and update the operand index. 982 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 983 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 984 MI->eraseFromParent(); 985 MI = MIB; 986 } 987 return MBB; 988 } 989 990 MachineBasicBlock * 991 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 992 MachineBasicBlock *MBB) const { 993 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 994 "Called emitXRayCustomEvent on the wrong MI!"); 995 auto &MF = *MI.getMF(); 996 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 997 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 998 MIB.add(MI.getOperand(OpIdx)); 999 1000 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1001 MI.eraseFromParent(); 1002 return MBB; 1003 } 1004 1005 MachineBasicBlock * 1006 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1007 MachineBasicBlock *MBB) const { 1008 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1009 "Called emitXRayTypedEvent on the wrong MI!"); 1010 auto &MF = *MI.getMF(); 1011 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1012 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1013 MIB.add(MI.getOperand(OpIdx)); 1014 1015 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1016 MI.eraseFromParent(); 1017 return MBB; 1018 } 1019 1020 /// findRepresentativeClass - Return the largest legal super-reg register class 1021 /// of the register class for the specified type and its associated "cost". 1022 // This function is in TargetLowering because it uses RegClassForVT which would 1023 // need to be moved to TargetRegisterInfo and would necessitate moving 1024 // isTypeLegal over as well - a massive change that would just require 1025 // TargetLowering having a TargetRegisterInfo class member that it would use. 1026 std::pair<const TargetRegisterClass *, uint8_t> 1027 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1028 MVT VT) const { 1029 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1030 if (!RC) 1031 return std::make_pair(RC, 0); 1032 1033 // Compute the set of all super-register classes. 1034 BitVector SuperRegRC(TRI->getNumRegClasses()); 1035 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1036 SuperRegRC.setBitsInMask(RCI.getMask()); 1037 1038 // Find the first legal register class with the largest spill size. 1039 const TargetRegisterClass *BestRC = RC; 1040 for (unsigned i : SuperRegRC.set_bits()) { 1041 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1042 // We want the largest possible spill size. 1043 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1044 continue; 1045 if (!isLegalRC(*TRI, *SuperRC)) 1046 continue; 1047 BestRC = SuperRC; 1048 } 1049 return std::make_pair(BestRC, 1); 1050 } 1051 1052 /// computeRegisterProperties - Once all of the register classes are added, 1053 /// this allows us to compute derived properties we expose. 1054 void TargetLoweringBase::computeRegisterProperties( 1055 const TargetRegisterInfo *TRI) { 1056 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1057 "Too many value types for ValueTypeActions to hold!"); 1058 1059 // Everything defaults to needing one register. 1060 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1061 NumRegistersForVT[i] = 1; 1062 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1063 } 1064 // ...except isVoid, which doesn't need any registers. 1065 NumRegistersForVT[MVT::isVoid] = 0; 1066 1067 // Find the largest integer register class. 1068 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1069 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1070 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1071 1072 // Every integer value type larger than this largest register takes twice as 1073 // many registers to represent as the previous ValueType. 1074 for (unsigned ExpandedReg = LargestIntReg + 1; 1075 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1076 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1077 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1078 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1079 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1080 TypeExpandInteger); 1081 } 1082 1083 // Inspect all of the ValueType's smaller than the largest integer 1084 // register to see which ones need promotion. 1085 unsigned LegalIntReg = LargestIntReg; 1086 for (unsigned IntReg = LargestIntReg - 1; 1087 IntReg >= (unsigned)MVT::i1; --IntReg) { 1088 MVT IVT = (MVT::SimpleValueType)IntReg; 1089 if (isTypeLegal(IVT)) { 1090 LegalIntReg = IntReg; 1091 } else { 1092 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1093 (const MVT::SimpleValueType)LegalIntReg; 1094 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1095 } 1096 } 1097 1098 // ppcf128 type is really two f64's. 1099 if (!isTypeLegal(MVT::ppcf128)) { 1100 if (isTypeLegal(MVT::f64)) { 1101 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1102 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1103 TransformToType[MVT::ppcf128] = MVT::f64; 1104 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1105 } else { 1106 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1107 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1108 TransformToType[MVT::ppcf128] = MVT::i128; 1109 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1110 } 1111 } 1112 1113 // Decide how to handle f128. If the target does not have native f128 support, 1114 // expand it to i128 and we will be generating soft float library calls. 1115 if (!isTypeLegal(MVT::f128)) { 1116 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1117 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1118 TransformToType[MVT::f128] = MVT::i128; 1119 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1120 } 1121 1122 // Decide how to handle f64. If the target does not have native f64 support, 1123 // expand it to i64 and we will be generating soft float library calls. 1124 if (!isTypeLegal(MVT::f64)) { 1125 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1126 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1127 TransformToType[MVT::f64] = MVT::i64; 1128 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1129 } 1130 1131 // Decide how to handle f32. If the target does not have native f32 support, 1132 // expand it to i32 and we will be generating soft float library calls. 1133 if (!isTypeLegal(MVT::f32)) { 1134 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1135 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1136 TransformToType[MVT::f32] = MVT::i32; 1137 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1138 } 1139 1140 // Decide how to handle f16. If the target does not have native f16 support, 1141 // promote it to f32, because there are no f16 library calls (except for 1142 // conversions). 1143 if (!isTypeLegal(MVT::f16)) { 1144 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1145 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1146 TransformToType[MVT::f16] = MVT::f32; 1147 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1148 } 1149 1150 // Loop over all of the vector value types to see which need transformations. 1151 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1152 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1153 MVT VT = (MVT::SimpleValueType) i; 1154 if (isTypeLegal(VT)) 1155 continue; 1156 1157 MVT EltVT = VT.getVectorElementType(); 1158 unsigned NElts = VT.getVectorNumElements(); 1159 bool IsLegalWiderType = false; 1160 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1161 switch (PreferredAction) { 1162 case TypePromoteInteger: 1163 // Try to promote the elements of integer vectors. If no legal 1164 // promotion was found, fall through to the widen-vector method. 1165 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1166 MVT SVT = (MVT::SimpleValueType) nVT; 1167 // Promote vectors of integers to vectors with the same number 1168 // of elements, with a wider element type. 1169 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1170 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1171 TransformToType[i] = SVT; 1172 RegisterTypeForVT[i] = SVT; 1173 NumRegistersForVT[i] = 1; 1174 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1175 IsLegalWiderType = true; 1176 break; 1177 } 1178 } 1179 if (IsLegalWiderType) 1180 break; 1181 LLVM_FALLTHROUGH; 1182 1183 case TypeWidenVector: 1184 // Try to widen the vector. 1185 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1186 MVT SVT = (MVT::SimpleValueType) nVT; 1187 if (SVT.getVectorElementType() == EltVT 1188 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1189 TransformToType[i] = SVT; 1190 RegisterTypeForVT[i] = SVT; 1191 NumRegistersForVT[i] = 1; 1192 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1193 IsLegalWiderType = true; 1194 break; 1195 } 1196 } 1197 if (IsLegalWiderType) 1198 break; 1199 LLVM_FALLTHROUGH; 1200 1201 case TypeSplitVector: 1202 case TypeScalarizeVector: { 1203 MVT IntermediateVT; 1204 MVT RegisterVT; 1205 unsigned NumIntermediates; 1206 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1207 NumIntermediates, RegisterVT, this); 1208 RegisterTypeForVT[i] = RegisterVT; 1209 1210 MVT NVT = VT.getPow2VectorType(); 1211 if (NVT == VT) { 1212 // Type is already a power of 2. The default action is to split. 1213 TransformToType[i] = MVT::Other; 1214 if (PreferredAction == TypeScalarizeVector) 1215 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1216 else if (PreferredAction == TypeSplitVector) 1217 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1218 else 1219 // Set type action according to the number of elements. 1220 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1221 : TypeSplitVector); 1222 } else { 1223 TransformToType[i] = NVT; 1224 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1225 } 1226 break; 1227 } 1228 default: 1229 llvm_unreachable("Unknown vector legalization action!"); 1230 } 1231 } 1232 1233 // Determine the 'representative' register class for each value type. 1234 // An representative register class is the largest (meaning one which is 1235 // not a sub-register class / subreg register class) legal register class for 1236 // a group of value types. For example, on i386, i8, i16, and i32 1237 // representative would be GR32; while on x86_64 it's GR64. 1238 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1239 const TargetRegisterClass* RRC; 1240 uint8_t Cost; 1241 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1242 RepRegClassForVT[i] = RRC; 1243 RepRegClassCostForVT[i] = Cost; 1244 } 1245 } 1246 1247 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1248 EVT VT) const { 1249 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1250 return getPointerTy(DL).SimpleTy; 1251 } 1252 1253 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1254 return MVT::i32; // return the default value 1255 } 1256 1257 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1258 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1259 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1260 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1261 /// 1262 /// This method returns the number of registers needed, and the VT for each 1263 /// register. It also returns the VT and quantity of the intermediate values 1264 /// before they are promoted/expanded. 1265 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1266 EVT &IntermediateVT, 1267 unsigned &NumIntermediates, 1268 MVT &RegisterVT) const { 1269 unsigned NumElts = VT.getVectorNumElements(); 1270 1271 // If there is a wider vector type with the same element type as this one, 1272 // or a promoted vector type that has the same number of elements which 1273 // are wider, then we should convert to that legal vector type. 1274 // This handles things like <2 x float> -> <4 x float> and 1275 // <4 x i1> -> <4 x i32>. 1276 LegalizeTypeAction TA = getTypeAction(Context, VT); 1277 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1278 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1279 if (isTypeLegal(RegisterEVT)) { 1280 IntermediateVT = RegisterEVT; 1281 RegisterVT = RegisterEVT.getSimpleVT(); 1282 NumIntermediates = 1; 1283 return 1; 1284 } 1285 } 1286 1287 // Figure out the right, legal destination reg to copy into. 1288 EVT EltTy = VT.getVectorElementType(); 1289 1290 unsigned NumVectorRegs = 1; 1291 1292 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1293 // could break down into LHS/RHS like LegalizeDAG does. 1294 if (!isPowerOf2_32(NumElts)) { 1295 NumVectorRegs = NumElts; 1296 NumElts = 1; 1297 } 1298 1299 // Divide the input until we get to a supported size. This will always 1300 // end with a scalar if the target doesn't support vectors. 1301 while (NumElts > 1 && !isTypeLegal( 1302 EVT::getVectorVT(Context, EltTy, NumElts))) { 1303 NumElts >>= 1; 1304 NumVectorRegs <<= 1; 1305 } 1306 1307 NumIntermediates = NumVectorRegs; 1308 1309 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1310 if (!isTypeLegal(NewVT)) 1311 NewVT = EltTy; 1312 IntermediateVT = NewVT; 1313 1314 MVT DestVT = getRegisterType(Context, NewVT); 1315 RegisterVT = DestVT; 1316 unsigned NewVTSize = NewVT.getSizeInBits(); 1317 1318 // Convert sizes such as i33 to i64. 1319 if (!isPowerOf2_32(NewVTSize)) 1320 NewVTSize = NextPowerOf2(NewVTSize); 1321 1322 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1323 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1324 1325 // Otherwise, promotion or legal types use the same number of registers as 1326 // the vector decimated to the appropriate level. 1327 return NumVectorRegs; 1328 } 1329 1330 /// Get the EVTs and ArgFlags collections that represent the legalized return 1331 /// type of the given function. This does not require a DAG or a return value, 1332 /// and is suitable for use before any DAGs for the function are constructed. 1333 /// TODO: Move this out of TargetLowering.cpp. 1334 void llvm::GetReturnInfo(Type *ReturnType, AttributeList attr, 1335 SmallVectorImpl<ISD::OutputArg> &Outs, 1336 const TargetLowering &TLI, const DataLayout &DL) { 1337 SmallVector<EVT, 4> ValueVTs; 1338 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1339 unsigned NumValues = ValueVTs.size(); 1340 if (NumValues == 0) return; 1341 1342 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1343 EVT VT = ValueVTs[j]; 1344 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1345 1346 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1347 ExtendKind = ISD::SIGN_EXTEND; 1348 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1349 ExtendKind = ISD::ZERO_EXTEND; 1350 1351 // FIXME: C calling convention requires the return type to be promoted to 1352 // at least 32-bit. But this is not necessary for non-C calling 1353 // conventions. The frontend should mark functions whose return values 1354 // require promoting with signext or zeroext attributes. 1355 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1356 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1357 if (VT.bitsLT(MinVT)) 1358 VT = MinVT; 1359 } 1360 1361 unsigned NumParts = 1362 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), VT); 1363 MVT PartVT = 1364 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), VT); 1365 1366 // 'inreg' on function refers to return value 1367 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1368 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1369 Flags.setInReg(); 1370 1371 // Propagate extension type if any 1372 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1373 Flags.setSExt(); 1374 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1375 Flags.setZExt(); 1376 1377 for (unsigned i = 0; i < NumParts; ++i) 1378 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0)); 1379 } 1380 } 1381 1382 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1383 /// function arguments in the caller parameter area. This is the actual 1384 /// alignment, not its logarithm. 1385 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1386 const DataLayout &DL) const { 1387 return DL.getABITypeAlignment(Ty); 1388 } 1389 1390 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1391 const DataLayout &DL, EVT VT, 1392 unsigned AddrSpace, 1393 unsigned Alignment, 1394 bool *Fast) const { 1395 // Check if the specified alignment is sufficient based on the data layout. 1396 // TODO: While using the data layout works in practice, a better solution 1397 // would be to implement this check directly (make this a virtual function). 1398 // For example, the ABI alignment may change based on software platform while 1399 // this function should only be affected by hardware implementation. 1400 Type *Ty = VT.getTypeForEVT(Context); 1401 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1402 // Assume that an access that meets the ABI-specified alignment is fast. 1403 if (Fast != nullptr) 1404 *Fast = true; 1405 return true; 1406 } 1407 1408 // This is a misaligned access. 1409 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast); 1410 } 1411 1412 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1413 return BranchProbability(MinPercentageForPredictableBranch, 100); 1414 } 1415 1416 //===----------------------------------------------------------------------===// 1417 // TargetTransformInfo Helpers 1418 //===----------------------------------------------------------------------===// 1419 1420 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1421 enum InstructionOpcodes { 1422 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1423 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1424 #include "llvm/IR/Instruction.def" 1425 }; 1426 switch (static_cast<InstructionOpcodes>(Opcode)) { 1427 case Ret: return 0; 1428 case Br: return 0; 1429 case Switch: return 0; 1430 case IndirectBr: return 0; 1431 case Invoke: return 0; 1432 case Resume: return 0; 1433 case Unreachable: return 0; 1434 case CleanupRet: return 0; 1435 case CatchRet: return 0; 1436 case CatchPad: return 0; 1437 case CatchSwitch: return 0; 1438 case CleanupPad: return 0; 1439 case Add: return ISD::ADD; 1440 case FAdd: return ISD::FADD; 1441 case Sub: return ISD::SUB; 1442 case FSub: return ISD::FSUB; 1443 case Mul: return ISD::MUL; 1444 case FMul: return ISD::FMUL; 1445 case UDiv: return ISD::UDIV; 1446 case SDiv: return ISD::SDIV; 1447 case FDiv: return ISD::FDIV; 1448 case URem: return ISD::UREM; 1449 case SRem: return ISD::SREM; 1450 case FRem: return ISD::FREM; 1451 case Shl: return ISD::SHL; 1452 case LShr: return ISD::SRL; 1453 case AShr: return ISD::SRA; 1454 case And: return ISD::AND; 1455 case Or: return ISD::OR; 1456 case Xor: return ISD::XOR; 1457 case Alloca: return 0; 1458 case Load: return ISD::LOAD; 1459 case Store: return ISD::STORE; 1460 case GetElementPtr: return 0; 1461 case Fence: return 0; 1462 case AtomicCmpXchg: return 0; 1463 case AtomicRMW: return 0; 1464 case Trunc: return ISD::TRUNCATE; 1465 case ZExt: return ISD::ZERO_EXTEND; 1466 case SExt: return ISD::SIGN_EXTEND; 1467 case FPToUI: return ISD::FP_TO_UINT; 1468 case FPToSI: return ISD::FP_TO_SINT; 1469 case UIToFP: return ISD::UINT_TO_FP; 1470 case SIToFP: return ISD::SINT_TO_FP; 1471 case FPTrunc: return ISD::FP_ROUND; 1472 case FPExt: return ISD::FP_EXTEND; 1473 case PtrToInt: return ISD::BITCAST; 1474 case IntToPtr: return ISD::BITCAST; 1475 case BitCast: return ISD::BITCAST; 1476 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1477 case ICmp: return ISD::SETCC; 1478 case FCmp: return ISD::SETCC; 1479 case PHI: return 0; 1480 case Call: return 0; 1481 case Select: return ISD::SELECT; 1482 case UserOp1: return 0; 1483 case UserOp2: return 0; 1484 case VAArg: return 0; 1485 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1486 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1487 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1488 case ExtractValue: return ISD::MERGE_VALUES; 1489 case InsertValue: return ISD::MERGE_VALUES; 1490 case LandingPad: return 0; 1491 } 1492 1493 llvm_unreachable("Unknown instruction type encountered!"); 1494 } 1495 1496 std::pair<int, MVT> 1497 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1498 Type *Ty) const { 1499 LLVMContext &C = Ty->getContext(); 1500 EVT MTy = getValueType(DL, Ty); 1501 1502 int Cost = 1; 1503 // We keep legalizing the type until we find a legal kind. We assume that 1504 // the only operation that costs anything is the split. After splitting 1505 // we need to handle two types. 1506 while (true) { 1507 LegalizeKind LK = getTypeConversion(C, MTy); 1508 1509 if (LK.first == TypeLegal) 1510 return std::make_pair(Cost, MTy.getSimpleVT()); 1511 1512 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1513 Cost *= 2; 1514 1515 // Do not loop with f128 type. 1516 if (MTy == LK.second) 1517 return std::make_pair(Cost, MTy.getSimpleVT()); 1518 1519 // Keep legalizing the type. 1520 MTy = LK.second; 1521 } 1522 } 1523 1524 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1525 bool UseTLS) const { 1526 // compiler-rt provides a variable with a magic name. Targets that do not 1527 // link with compiler-rt may also provide such a variable. 1528 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1529 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1530 auto UnsafeStackPtr = 1531 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1532 1533 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1534 1535 if (!UnsafeStackPtr) { 1536 auto TLSModel = UseTLS ? 1537 GlobalValue::InitialExecTLSModel : 1538 GlobalValue::NotThreadLocal; 1539 // The global variable is not defined yet, define it ourselves. 1540 // We use the initial-exec TLS model because we do not support the 1541 // variable living anywhere other than in the main executable. 1542 UnsafeStackPtr = new GlobalVariable( 1543 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1544 UnsafeStackPtrVar, nullptr, TLSModel); 1545 } else { 1546 // The variable exists, check its type and attributes. 1547 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1548 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1549 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1550 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1551 (UseTLS ? "" : "not ") + "be thread-local"); 1552 } 1553 return UnsafeStackPtr; 1554 } 1555 1556 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1557 if (!TM.getTargetTriple().isAndroid()) 1558 return getDefaultSafeStackPointerLocation(IRB, true); 1559 1560 // Android provides a libc function to retrieve the address of the current 1561 // thread's unsafe stack pointer. 1562 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1563 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1564 Value *Fn = M->getOrInsertFunction("__safestack_pointer_address", 1565 StackPtrTy->getPointerTo(0)); 1566 return IRB.CreateCall(Fn); 1567 } 1568 1569 //===----------------------------------------------------------------------===// 1570 // Loop Strength Reduction hooks 1571 //===----------------------------------------------------------------------===// 1572 1573 /// isLegalAddressingMode - Return true if the addressing mode represented 1574 /// by AM is legal for this target, for a load/store of the specified type. 1575 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1576 const AddrMode &AM, Type *Ty, 1577 unsigned AS, Instruction *I) const { 1578 // The default implementation of this implements a conservative RISCy, r+r and 1579 // r+i addr mode. 1580 1581 // Allows a sign-extended 16-bit immediate field. 1582 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1583 return false; 1584 1585 // No global is ever allowed as a base. 1586 if (AM.BaseGV) 1587 return false; 1588 1589 // Only support r+r, 1590 switch (AM.Scale) { 1591 case 0: // "r+i" or just "i", depending on HasBaseReg. 1592 break; 1593 case 1: 1594 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1595 return false; 1596 // Otherwise we have r+r or r+i. 1597 break; 1598 case 2: 1599 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1600 return false; 1601 // Allow 2*r as r+r. 1602 break; 1603 default: // Don't allow n * r 1604 return false; 1605 } 1606 1607 return true; 1608 } 1609 1610 //===----------------------------------------------------------------------===// 1611 // Stack Protector 1612 //===----------------------------------------------------------------------===// 1613 1614 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1615 // so that SelectionDAG handle SSP. 1616 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1617 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1618 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1619 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1620 return M.getOrInsertGlobal("__guard_local", PtrTy); 1621 } 1622 return nullptr; 1623 } 1624 1625 // Currently only support "standard" __stack_chk_guard. 1626 // TODO: add LOAD_STACK_GUARD support. 1627 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1628 if (!M.getNamedValue("__stack_chk_guard")) 1629 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1630 GlobalVariable::ExternalLinkage, 1631 nullptr, "__stack_chk_guard"); 1632 } 1633 1634 // Currently only support "standard" __stack_chk_guard. 1635 // TODO: add LOAD_STACK_GUARD support. 1636 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1637 return M.getNamedValue("__stack_chk_guard"); 1638 } 1639 1640 Value *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1641 return nullptr; 1642 } 1643 1644 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1645 return MinimumJumpTableEntries; 1646 } 1647 1648 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1649 MinimumJumpTableEntries = Val; 1650 } 1651 1652 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1653 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1654 } 1655 1656 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1657 return MaximumJumpTableSize; 1658 } 1659 1660 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1661 MaximumJumpTableSize = Val; 1662 } 1663 1664 //===----------------------------------------------------------------------===// 1665 // Reciprocal Estimates 1666 //===----------------------------------------------------------------------===// 1667 1668 /// Get the reciprocal estimate attribute string for a function that will 1669 /// override the target defaults. 1670 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1671 const Function &F = MF.getFunction(); 1672 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1673 } 1674 1675 /// Construct a string for the given reciprocal operation of the given type. 1676 /// This string should match the corresponding option to the front-end's 1677 /// "-mrecip" flag assuming those strings have been passed through in an 1678 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1679 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1680 std::string Name = VT.isVector() ? "vec-" : ""; 1681 1682 Name += IsSqrt ? "sqrt" : "div"; 1683 1684 // TODO: Handle "half" or other float types? 1685 if (VT.getScalarType() == MVT::f64) { 1686 Name += "d"; 1687 } else { 1688 assert(VT.getScalarType() == MVT::f32 && 1689 "Unexpected FP type for reciprocal estimate"); 1690 Name += "f"; 1691 } 1692 1693 return Name; 1694 } 1695 1696 /// Return the character position and value (a single numeric character) of a 1697 /// customized refinement operation in the input string if it exists. Return 1698 /// false if there is no customized refinement step count. 1699 static bool parseRefinementStep(StringRef In, size_t &Position, 1700 uint8_t &Value) { 1701 const char RefStepToken = ':'; 1702 Position = In.find(RefStepToken); 1703 if (Position == StringRef::npos) 1704 return false; 1705 1706 StringRef RefStepString = In.substr(Position + 1); 1707 // Allow exactly one numeric character for the additional refinement 1708 // step parameter. 1709 if (RefStepString.size() == 1) { 1710 char RefStepChar = RefStepString[0]; 1711 if (RefStepChar >= '0' && RefStepChar <= '9') { 1712 Value = RefStepChar - '0'; 1713 return true; 1714 } 1715 } 1716 report_fatal_error("Invalid refinement step for -recip."); 1717 } 1718 1719 /// For the input attribute string, return one of the ReciprocalEstimate enum 1720 /// status values (enabled, disabled, or not specified) for this operation on 1721 /// the specified data type. 1722 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1723 if (Override.empty()) 1724 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1725 1726 SmallVector<StringRef, 4> OverrideVector; 1727 Override.split(OverrideVector, ','); 1728 unsigned NumArgs = OverrideVector.size(); 1729 1730 // Check if "all", "none", or "default" was specified. 1731 if (NumArgs == 1) { 1732 // Look for an optional setting of the number of refinement steps needed 1733 // for this type of reciprocal operation. 1734 size_t RefPos; 1735 uint8_t RefSteps; 1736 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1737 // Split the string for further processing. 1738 Override = Override.substr(0, RefPos); 1739 } 1740 1741 // All reciprocal types are enabled. 1742 if (Override == "all") 1743 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1744 1745 // All reciprocal types are disabled. 1746 if (Override == "none") 1747 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1748 1749 // Target defaults for enablement are used. 1750 if (Override == "default") 1751 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1752 } 1753 1754 // The attribute string may omit the size suffix ('f'/'d'). 1755 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1756 std::string VTNameNoSize = VTName; 1757 VTNameNoSize.pop_back(); 1758 static const char DisabledPrefix = '!'; 1759 1760 for (StringRef RecipType : OverrideVector) { 1761 size_t RefPos; 1762 uint8_t RefSteps; 1763 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1764 RecipType = RecipType.substr(0, RefPos); 1765 1766 // Ignore the disablement token for string matching. 1767 bool IsDisabled = RecipType[0] == DisabledPrefix; 1768 if (IsDisabled) 1769 RecipType = RecipType.substr(1); 1770 1771 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1772 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1773 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1774 } 1775 1776 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1777 } 1778 1779 /// For the input attribute string, return the customized refinement step count 1780 /// for this operation on the specified data type. If the step count does not 1781 /// exist, return the ReciprocalEstimate enum value for unspecified. 1782 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1783 if (Override.empty()) 1784 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1785 1786 SmallVector<StringRef, 4> OverrideVector; 1787 Override.split(OverrideVector, ','); 1788 unsigned NumArgs = OverrideVector.size(); 1789 1790 // Check if "all", "default", or "none" was specified. 1791 if (NumArgs == 1) { 1792 // Look for an optional setting of the number of refinement steps needed 1793 // for this type of reciprocal operation. 1794 size_t RefPos; 1795 uint8_t RefSteps; 1796 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1797 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1798 1799 // Split the string for further processing. 1800 Override = Override.substr(0, RefPos); 1801 assert(Override != "none" && 1802 "Disabled reciprocals, but specifed refinement steps?"); 1803 1804 // If this is a general override, return the specified number of steps. 1805 if (Override == "all" || Override == "default") 1806 return RefSteps; 1807 } 1808 1809 // The attribute string may omit the size suffix ('f'/'d'). 1810 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1811 std::string VTNameNoSize = VTName; 1812 VTNameNoSize.pop_back(); 1813 1814 for (StringRef RecipType : OverrideVector) { 1815 size_t RefPos; 1816 uint8_t RefSteps; 1817 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1818 continue; 1819 1820 RecipType = RecipType.substr(0, RefPos); 1821 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1822 return RefSteps; 1823 } 1824 1825 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1826 } 1827 1828 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1829 MachineFunction &MF) const { 1830 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1831 } 1832 1833 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1834 MachineFunction &MF) const { 1835 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1836 } 1837 1838 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1839 MachineFunction &MF) const { 1840 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1841 } 1842 1843 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1844 MachineFunction &MF) const { 1845 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1846 } 1847 1848 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1849 MF.getRegInfo().freezeReservedRegs(MF); 1850 } 1851