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/IR/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 (TT.isAndroid() && !TT.isAndroidVersionLT(9))) { 166 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 167 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 168 setLibcallName(RTLIB::SINCOS_F80, "sincosl"); 169 setLibcallName(RTLIB::SINCOS_F128, "sincosl"); 170 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl"); 171 } 172 173 if (TT.isOSOpenBSD()) { 174 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr); 175 } 176 } 177 178 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 179 /// UNKNOWN_LIBCALL if there is none. 180 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 181 if (OpVT == MVT::f16) { 182 if (RetVT == MVT::f32) 183 return FPEXT_F16_F32; 184 } else if (OpVT == MVT::f32) { 185 if (RetVT == MVT::f64) 186 return FPEXT_F32_F64; 187 if (RetVT == MVT::f128) 188 return FPEXT_F32_F128; 189 if (RetVT == MVT::ppcf128) 190 return FPEXT_F32_PPCF128; 191 } else if (OpVT == MVT::f64) { 192 if (RetVT == MVT::f128) 193 return FPEXT_F64_F128; 194 else if (RetVT == MVT::ppcf128) 195 return FPEXT_F64_PPCF128; 196 } else if (OpVT == MVT::f80) { 197 if (RetVT == MVT::f128) 198 return FPEXT_F80_F128; 199 } 200 201 return UNKNOWN_LIBCALL; 202 } 203 204 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 205 /// UNKNOWN_LIBCALL if there is none. 206 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 207 if (RetVT == MVT::f16) { 208 if (OpVT == MVT::f32) 209 return FPROUND_F32_F16; 210 if (OpVT == MVT::f64) 211 return FPROUND_F64_F16; 212 if (OpVT == MVT::f80) 213 return FPROUND_F80_F16; 214 if (OpVT == MVT::f128) 215 return FPROUND_F128_F16; 216 if (OpVT == MVT::ppcf128) 217 return FPROUND_PPCF128_F16; 218 } else if (RetVT == MVT::f32) { 219 if (OpVT == MVT::f64) 220 return FPROUND_F64_F32; 221 if (OpVT == MVT::f80) 222 return FPROUND_F80_F32; 223 if (OpVT == MVT::f128) 224 return FPROUND_F128_F32; 225 if (OpVT == MVT::ppcf128) 226 return FPROUND_PPCF128_F32; 227 } else if (RetVT == MVT::f64) { 228 if (OpVT == MVT::f80) 229 return FPROUND_F80_F64; 230 if (OpVT == MVT::f128) 231 return FPROUND_F128_F64; 232 if (OpVT == MVT::ppcf128) 233 return FPROUND_PPCF128_F64; 234 } else if (RetVT == MVT::f80) { 235 if (OpVT == MVT::f128) 236 return FPROUND_F128_F80; 237 } 238 239 return UNKNOWN_LIBCALL; 240 } 241 242 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 243 /// UNKNOWN_LIBCALL if there is none. 244 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 245 if (OpVT == MVT::f32) { 246 if (RetVT == MVT::i32) 247 return FPTOSINT_F32_I32; 248 if (RetVT == MVT::i64) 249 return FPTOSINT_F32_I64; 250 if (RetVT == MVT::i128) 251 return FPTOSINT_F32_I128; 252 } else if (OpVT == MVT::f64) { 253 if (RetVT == MVT::i32) 254 return FPTOSINT_F64_I32; 255 if (RetVT == MVT::i64) 256 return FPTOSINT_F64_I64; 257 if (RetVT == MVT::i128) 258 return FPTOSINT_F64_I128; 259 } else if (OpVT == MVT::f80) { 260 if (RetVT == MVT::i32) 261 return FPTOSINT_F80_I32; 262 if (RetVT == MVT::i64) 263 return FPTOSINT_F80_I64; 264 if (RetVT == MVT::i128) 265 return FPTOSINT_F80_I128; 266 } else if (OpVT == MVT::f128) { 267 if (RetVT == MVT::i32) 268 return FPTOSINT_F128_I32; 269 if (RetVT == MVT::i64) 270 return FPTOSINT_F128_I64; 271 if (RetVT == MVT::i128) 272 return FPTOSINT_F128_I128; 273 } else if (OpVT == MVT::ppcf128) { 274 if (RetVT == MVT::i32) 275 return FPTOSINT_PPCF128_I32; 276 if (RetVT == MVT::i64) 277 return FPTOSINT_PPCF128_I64; 278 if (RetVT == MVT::i128) 279 return FPTOSINT_PPCF128_I128; 280 } 281 return UNKNOWN_LIBCALL; 282 } 283 284 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 285 /// UNKNOWN_LIBCALL if there is none. 286 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 287 if (OpVT == MVT::f32) { 288 if (RetVT == MVT::i32) 289 return FPTOUINT_F32_I32; 290 if (RetVT == MVT::i64) 291 return FPTOUINT_F32_I64; 292 if (RetVT == MVT::i128) 293 return FPTOUINT_F32_I128; 294 } else if (OpVT == MVT::f64) { 295 if (RetVT == MVT::i32) 296 return FPTOUINT_F64_I32; 297 if (RetVT == MVT::i64) 298 return FPTOUINT_F64_I64; 299 if (RetVT == MVT::i128) 300 return FPTOUINT_F64_I128; 301 } else if (OpVT == MVT::f80) { 302 if (RetVT == MVT::i32) 303 return FPTOUINT_F80_I32; 304 if (RetVT == MVT::i64) 305 return FPTOUINT_F80_I64; 306 if (RetVT == MVT::i128) 307 return FPTOUINT_F80_I128; 308 } else if (OpVT == MVT::f128) { 309 if (RetVT == MVT::i32) 310 return FPTOUINT_F128_I32; 311 if (RetVT == MVT::i64) 312 return FPTOUINT_F128_I64; 313 if (RetVT == MVT::i128) 314 return FPTOUINT_F128_I128; 315 } else if (OpVT == MVT::ppcf128) { 316 if (RetVT == MVT::i32) 317 return FPTOUINT_PPCF128_I32; 318 if (RetVT == MVT::i64) 319 return FPTOUINT_PPCF128_I64; 320 if (RetVT == MVT::i128) 321 return FPTOUINT_PPCF128_I128; 322 } 323 return UNKNOWN_LIBCALL; 324 } 325 326 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 327 /// UNKNOWN_LIBCALL if there is none. 328 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 329 if (OpVT == MVT::i32) { 330 if (RetVT == MVT::f32) 331 return SINTTOFP_I32_F32; 332 if (RetVT == MVT::f64) 333 return SINTTOFP_I32_F64; 334 if (RetVT == MVT::f80) 335 return SINTTOFP_I32_F80; 336 if (RetVT == MVT::f128) 337 return SINTTOFP_I32_F128; 338 if (RetVT == MVT::ppcf128) 339 return SINTTOFP_I32_PPCF128; 340 } else if (OpVT == MVT::i64) { 341 if (RetVT == MVT::f32) 342 return SINTTOFP_I64_F32; 343 if (RetVT == MVT::f64) 344 return SINTTOFP_I64_F64; 345 if (RetVT == MVT::f80) 346 return SINTTOFP_I64_F80; 347 if (RetVT == MVT::f128) 348 return SINTTOFP_I64_F128; 349 if (RetVT == MVT::ppcf128) 350 return SINTTOFP_I64_PPCF128; 351 } else if (OpVT == MVT::i128) { 352 if (RetVT == MVT::f32) 353 return SINTTOFP_I128_F32; 354 if (RetVT == MVT::f64) 355 return SINTTOFP_I128_F64; 356 if (RetVT == MVT::f80) 357 return SINTTOFP_I128_F80; 358 if (RetVT == MVT::f128) 359 return SINTTOFP_I128_F128; 360 if (RetVT == MVT::ppcf128) 361 return SINTTOFP_I128_PPCF128; 362 } 363 return UNKNOWN_LIBCALL; 364 } 365 366 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 367 /// UNKNOWN_LIBCALL if there is none. 368 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 369 if (OpVT == MVT::i32) { 370 if (RetVT == MVT::f32) 371 return UINTTOFP_I32_F32; 372 if (RetVT == MVT::f64) 373 return UINTTOFP_I32_F64; 374 if (RetVT == MVT::f80) 375 return UINTTOFP_I32_F80; 376 if (RetVT == MVT::f128) 377 return UINTTOFP_I32_F128; 378 if (RetVT == MVT::ppcf128) 379 return UINTTOFP_I32_PPCF128; 380 } else if (OpVT == MVT::i64) { 381 if (RetVT == MVT::f32) 382 return UINTTOFP_I64_F32; 383 if (RetVT == MVT::f64) 384 return UINTTOFP_I64_F64; 385 if (RetVT == MVT::f80) 386 return UINTTOFP_I64_F80; 387 if (RetVT == MVT::f128) 388 return UINTTOFP_I64_F128; 389 if (RetVT == MVT::ppcf128) 390 return UINTTOFP_I64_PPCF128; 391 } else if (OpVT == MVT::i128) { 392 if (RetVT == MVT::f32) 393 return UINTTOFP_I128_F32; 394 if (RetVT == MVT::f64) 395 return UINTTOFP_I128_F64; 396 if (RetVT == MVT::f80) 397 return UINTTOFP_I128_F80; 398 if (RetVT == MVT::f128) 399 return UINTTOFP_I128_F128; 400 if (RetVT == MVT::ppcf128) 401 return UINTTOFP_I128_PPCF128; 402 } 403 return UNKNOWN_LIBCALL; 404 } 405 406 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 407 #define OP_TO_LIBCALL(Name, Enum) \ 408 case Name: \ 409 switch (VT.SimpleTy) { \ 410 default: \ 411 return UNKNOWN_LIBCALL; \ 412 case MVT::i8: \ 413 return Enum##_1; \ 414 case MVT::i16: \ 415 return Enum##_2; \ 416 case MVT::i32: \ 417 return Enum##_4; \ 418 case MVT::i64: \ 419 return Enum##_8; \ 420 case MVT::i128: \ 421 return Enum##_16; \ 422 } 423 424 switch (Opc) { 425 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 426 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 427 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 428 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 429 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 430 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 431 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 432 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 433 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 434 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 435 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 436 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 437 } 438 439 #undef OP_TO_LIBCALL 440 441 return UNKNOWN_LIBCALL; 442 } 443 444 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 445 switch (ElementSize) { 446 case 1: 447 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 448 case 2: 449 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 450 case 4: 451 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 452 case 8: 453 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 454 case 16: 455 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 456 default: 457 return UNKNOWN_LIBCALL; 458 } 459 } 460 461 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 462 switch (ElementSize) { 463 case 1: 464 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 465 case 2: 466 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 467 case 4: 468 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 469 case 8: 470 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 471 case 16: 472 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 473 default: 474 return UNKNOWN_LIBCALL; 475 } 476 } 477 478 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 479 switch (ElementSize) { 480 case 1: 481 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 482 case 2: 483 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 484 case 4: 485 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 486 case 8: 487 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 488 case 16: 489 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 490 default: 491 return UNKNOWN_LIBCALL; 492 } 493 } 494 495 /// InitCmpLibcallCCs - Set default comparison libcall CC. 496 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 497 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 498 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 499 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 500 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 501 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 502 CCs[RTLIB::UNE_F32] = ISD::SETNE; 503 CCs[RTLIB::UNE_F64] = ISD::SETNE; 504 CCs[RTLIB::UNE_F128] = ISD::SETNE; 505 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 506 CCs[RTLIB::OGE_F32] = ISD::SETGE; 507 CCs[RTLIB::OGE_F64] = ISD::SETGE; 508 CCs[RTLIB::OGE_F128] = ISD::SETGE; 509 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 510 CCs[RTLIB::OLT_F32] = ISD::SETLT; 511 CCs[RTLIB::OLT_F64] = ISD::SETLT; 512 CCs[RTLIB::OLT_F128] = ISD::SETLT; 513 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 514 CCs[RTLIB::OLE_F32] = ISD::SETLE; 515 CCs[RTLIB::OLE_F64] = ISD::SETLE; 516 CCs[RTLIB::OLE_F128] = ISD::SETLE; 517 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 518 CCs[RTLIB::OGT_F32] = ISD::SETGT; 519 CCs[RTLIB::OGT_F64] = ISD::SETGT; 520 CCs[RTLIB::OGT_F128] = ISD::SETGT; 521 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 522 CCs[RTLIB::UO_F32] = ISD::SETNE; 523 CCs[RTLIB::UO_F64] = ISD::SETNE; 524 CCs[RTLIB::UO_F128] = ISD::SETNE; 525 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 526 CCs[RTLIB::O_F32] = ISD::SETEQ; 527 CCs[RTLIB::O_F64] = ISD::SETEQ; 528 CCs[RTLIB::O_F128] = ISD::SETEQ; 529 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 530 } 531 532 /// NOTE: The TargetMachine owns TLOF. 533 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 534 initActions(); 535 536 // Perform these initializations only once. 537 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 538 MaxLoadsPerMemcmp = 8; 539 MaxGluedStoresPerMemcpy = 0; 540 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 541 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 542 UseUnderscoreSetJmp = false; 543 UseUnderscoreLongJmp = false; 544 HasMultipleConditionRegisters = false; 545 HasExtractBitsInsn = false; 546 JumpIsExpensive = JumpIsExpensiveOverride; 547 PredictableSelectIsExpensive = false; 548 EnableExtLdPromotion = false; 549 HasFloatingPointExceptions = true; 550 StackPointerRegisterToSaveRestore = 0; 551 BooleanContents = UndefinedBooleanContent; 552 BooleanFloatContents = UndefinedBooleanContent; 553 BooleanVectorContents = UndefinedBooleanContent; 554 SchedPreferenceInfo = Sched::ILP; 555 JumpBufSize = 0; 556 JumpBufAlignment = 0; 557 MinFunctionAlignment = 0; 558 PrefFunctionAlignment = 0; 559 PrefLoopAlignment = 0; 560 GatherAllAliasesMaxDepth = 18; 561 MinStackArgumentAlignment = 1; 562 // TODO: the default will be switched to 0 in the next commit, along 563 // with the Target-specific changes necessary. 564 MaxAtomicSizeInBitsSupported = 1024; 565 566 MinCmpXchgSizeInBits = 0; 567 SupportsUnalignedAtomics = false; 568 569 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 570 571 InitLibcalls(TM.getTargetTriple()); 572 InitCmpLibcallCCs(CmpLibcallCCs); 573 } 574 575 void TargetLoweringBase::initActions() { 576 // All operations default to being supported. 577 memset(OpActions, 0, sizeof(OpActions)); 578 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 579 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 580 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 581 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 582 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 583 std::fill(std::begin(TargetDAGCombineArray), 584 std::end(TargetDAGCombineArray), 0); 585 586 // Set default actions for various operations. 587 for (MVT VT : MVT::all_valuetypes()) { 588 // Default all indexed load / store to expand. 589 for (unsigned IM = (unsigned)ISD::PRE_INC; 590 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 591 setIndexedLoadAction(IM, VT, Expand); 592 setIndexedStoreAction(IM, VT, Expand); 593 } 594 595 // Most backends expect to see the node which just returns the value loaded. 596 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 597 598 // These operations default to expand. 599 setOperationAction(ISD::FGETSIGN, VT, Expand); 600 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 601 setOperationAction(ISD::FMINNUM, VT, Expand); 602 setOperationAction(ISD::FMAXNUM, VT, Expand); 603 setOperationAction(ISD::FMINNUM_IEEE, VT, Expand); 604 setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand); 605 setOperationAction(ISD::FMINIMUM, VT, Expand); 606 setOperationAction(ISD::FMAXIMUM, VT, Expand); 607 setOperationAction(ISD::FMAD, VT, Expand); 608 setOperationAction(ISD::SMIN, VT, Expand); 609 setOperationAction(ISD::SMAX, VT, Expand); 610 setOperationAction(ISD::UMIN, VT, Expand); 611 setOperationAction(ISD::UMAX, VT, Expand); 612 setOperationAction(ISD::ABS, VT, Expand); 613 setOperationAction(ISD::FSHL, VT, Expand); 614 setOperationAction(ISD::FSHR, VT, Expand); 615 setOperationAction(ISD::SADDSAT, VT, Expand); 616 setOperationAction(ISD::UADDSAT, VT, Expand); 617 setOperationAction(ISD::SSUBSAT, VT, Expand); 618 setOperationAction(ISD::USUBSAT, VT, Expand); 619 setOperationAction(ISD::SMULFIX, VT, Expand); 620 621 // Overflow operations default to expand 622 setOperationAction(ISD::SADDO, VT, Expand); 623 setOperationAction(ISD::SSUBO, VT, Expand); 624 setOperationAction(ISD::UADDO, VT, Expand); 625 setOperationAction(ISD::USUBO, VT, Expand); 626 setOperationAction(ISD::SMULO, VT, Expand); 627 setOperationAction(ISD::UMULO, VT, Expand); 628 629 // ADDCARRY operations default to expand 630 setOperationAction(ISD::ADDCARRY, VT, Expand); 631 setOperationAction(ISD::SUBCARRY, VT, Expand); 632 setOperationAction(ISD::SETCCCARRY, VT, Expand); 633 634 // ADDC/ADDE/SUBC/SUBE default to expand. 635 setOperationAction(ISD::ADDC, VT, Expand); 636 setOperationAction(ISD::ADDE, VT, Expand); 637 setOperationAction(ISD::SUBC, VT, Expand); 638 setOperationAction(ISD::SUBE, VT, Expand); 639 640 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 641 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 642 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 643 644 setOperationAction(ISD::BITREVERSE, VT, Expand); 645 646 // These library functions default to expand. 647 setOperationAction(ISD::FROUND, VT, Expand); 648 setOperationAction(ISD::FPOWI, VT, Expand); 649 650 // These operations default to expand for vector types. 651 if (VT.isVector()) { 652 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 653 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 654 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 655 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 656 } 657 658 // For most targets @llvm.get.dynamic.area.offset just returns 0. 659 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 660 } 661 662 // Most targets ignore the @llvm.prefetch intrinsic. 663 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 664 665 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 666 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 667 668 // ConstantFP nodes default to expand. Targets can either change this to 669 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 670 // to optimize expansions for certain constants. 671 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 672 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 673 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 674 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 675 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 676 677 // These library functions default to expand. 678 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 679 setOperationAction(ISD::FCBRT, VT, Expand); 680 setOperationAction(ISD::FLOG , VT, Expand); 681 setOperationAction(ISD::FLOG2, VT, Expand); 682 setOperationAction(ISD::FLOG10, VT, Expand); 683 setOperationAction(ISD::FEXP , VT, Expand); 684 setOperationAction(ISD::FEXP2, VT, Expand); 685 setOperationAction(ISD::FFLOOR, VT, Expand); 686 setOperationAction(ISD::FNEARBYINT, VT, Expand); 687 setOperationAction(ISD::FCEIL, VT, Expand); 688 setOperationAction(ISD::FRINT, VT, Expand); 689 setOperationAction(ISD::FTRUNC, VT, Expand); 690 setOperationAction(ISD::FROUND, VT, Expand); 691 } 692 693 // Default ISD::TRAP to expand (which turns it into abort). 694 setOperationAction(ISD::TRAP, MVT::Other, Expand); 695 696 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 697 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 698 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 699 } 700 701 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 702 EVT) const { 703 return MVT::getIntegerVT(8 * DL.getPointerSize(0)); 704 } 705 706 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 707 bool LegalTypes) const { 708 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 709 if (LHSTy.isVector()) 710 return LHSTy; 711 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 712 : getPointerTy(DL); 713 } 714 715 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 716 assert(isTypeLegal(VT)); 717 switch (Op) { 718 default: 719 return false; 720 case ISD::SDIV: 721 case ISD::UDIV: 722 case ISD::SREM: 723 case ISD::UREM: 724 return true; 725 } 726 } 727 728 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 729 // If the command-line option was specified, ignore this request. 730 if (!JumpIsExpensiveOverride.getNumOccurrences()) 731 JumpIsExpensive = isExpensive; 732 } 733 734 TargetLoweringBase::LegalizeKind 735 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 736 // If this is a simple type, use the ComputeRegisterProp mechanism. 737 if (VT.isSimple()) { 738 MVT SVT = VT.getSimpleVT(); 739 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 740 MVT NVT = TransformToType[SVT.SimpleTy]; 741 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 742 743 assert((LA == TypeLegal || LA == TypeSoftenFloat || 744 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) && 745 "Promote may not follow Expand or Promote"); 746 747 if (LA == TypeSplitVector) 748 return LegalizeKind(LA, 749 EVT::getVectorVT(Context, SVT.getVectorElementType(), 750 SVT.getVectorNumElements() / 2)); 751 if (LA == TypeScalarizeVector) 752 return LegalizeKind(LA, SVT.getVectorElementType()); 753 return LegalizeKind(LA, NVT); 754 } 755 756 // Handle Extended Scalar Types. 757 if (!VT.isVector()) { 758 assert(VT.isInteger() && "Float types must be simple"); 759 unsigned BitSize = VT.getSizeInBits(); 760 // First promote to a power-of-two size, then expand if necessary. 761 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 762 EVT NVT = VT.getRoundIntegerType(Context); 763 assert(NVT != VT && "Unable to round integer VT"); 764 LegalizeKind NextStep = getTypeConversion(Context, NVT); 765 // Avoid multi-step promotion. 766 if (NextStep.first == TypePromoteInteger) 767 return NextStep; 768 // Return rounded integer type. 769 return LegalizeKind(TypePromoteInteger, NVT); 770 } 771 772 return LegalizeKind(TypeExpandInteger, 773 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 774 } 775 776 // Handle vector types. 777 unsigned NumElts = VT.getVectorNumElements(); 778 EVT EltVT = VT.getVectorElementType(); 779 780 // Vectors with only one element are always scalarized. 781 if (NumElts == 1) 782 return LegalizeKind(TypeScalarizeVector, EltVT); 783 784 // Try to widen vector elements until the element type is a power of two and 785 // promote it to a legal type later on, for example: 786 // <3 x i8> -> <4 x i8> -> <4 x i32> 787 if (EltVT.isInteger()) { 788 // Vectors with a number of elements that is not a power of two are always 789 // widened, for example <3 x i8> -> <4 x i8>. 790 if (!VT.isPow2VectorType()) { 791 NumElts = (unsigned)NextPowerOf2(NumElts); 792 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 793 return LegalizeKind(TypeWidenVector, NVT); 794 } 795 796 // Examine the element type. 797 LegalizeKind LK = getTypeConversion(Context, EltVT); 798 799 // If type is to be expanded, split the vector. 800 // <4 x i140> -> <2 x i140> 801 if (LK.first == TypeExpandInteger) 802 return LegalizeKind(TypeSplitVector, 803 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 804 805 // Promote the integer element types until a legal vector type is found 806 // or until the element integer type is too big. If a legal type was not 807 // found, fallback to the usual mechanism of widening/splitting the 808 // vector. 809 EVT OldEltVT = EltVT; 810 while (true) { 811 // Increase the bitwidth of the element to the next pow-of-two 812 // (which is greater than 8 bits). 813 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 814 .getRoundIntegerType(Context); 815 816 // Stop trying when getting a non-simple element type. 817 // Note that vector elements may be greater than legal vector element 818 // types. Example: X86 XMM registers hold 64bit element on 32bit 819 // systems. 820 if (!EltVT.isSimple()) 821 break; 822 823 // Build a new vector type and check if it is legal. 824 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 825 // Found a legal promoted vector type. 826 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 827 return LegalizeKind(TypePromoteInteger, 828 EVT::getVectorVT(Context, EltVT, NumElts)); 829 } 830 831 // Reset the type to the unexpanded type if we did not find a legal vector 832 // type with a promoted vector element type. 833 EltVT = OldEltVT; 834 } 835 836 // Try to widen the vector until a legal type is found. 837 // If there is no wider legal type, split the vector. 838 while (true) { 839 // Round up to the next power of 2. 840 NumElts = (unsigned)NextPowerOf2(NumElts); 841 842 // If there is no simple vector type with this many elements then there 843 // cannot be a larger legal vector type. Note that this assumes that 844 // there are no skipped intermediate vector types in the simple types. 845 if (!EltVT.isSimple()) 846 break; 847 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 848 if (LargerVector == MVT()) 849 break; 850 851 // If this type is legal then widen the vector. 852 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 853 return LegalizeKind(TypeWidenVector, LargerVector); 854 } 855 856 // Widen odd vectors to next power of two. 857 if (!VT.isPow2VectorType()) { 858 EVT NVT = VT.getPow2VectorType(Context); 859 return LegalizeKind(TypeWidenVector, NVT); 860 } 861 862 // Vectors with illegal element types are expanded. 863 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 864 return LegalizeKind(TypeSplitVector, NVT); 865 } 866 867 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 868 unsigned &NumIntermediates, 869 MVT &RegisterVT, 870 TargetLoweringBase *TLI) { 871 // Figure out the right, legal destination reg to copy into. 872 unsigned NumElts = VT.getVectorNumElements(); 873 MVT EltTy = VT.getVectorElementType(); 874 875 unsigned NumVectorRegs = 1; 876 877 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 878 // could break down into LHS/RHS like LegalizeDAG does. 879 if (!isPowerOf2_32(NumElts)) { 880 NumVectorRegs = NumElts; 881 NumElts = 1; 882 } 883 884 // Divide the input until we get to a supported size. This will always 885 // end with a scalar if the target doesn't support vectors. 886 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 887 NumElts >>= 1; 888 NumVectorRegs <<= 1; 889 } 890 891 NumIntermediates = NumVectorRegs; 892 893 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 894 if (!TLI->isTypeLegal(NewVT)) 895 NewVT = EltTy; 896 IntermediateVT = NewVT; 897 898 unsigned NewVTSize = NewVT.getSizeInBits(); 899 900 // Convert sizes such as i33 to i64. 901 if (!isPowerOf2_32(NewVTSize)) 902 NewVTSize = NextPowerOf2(NewVTSize); 903 904 MVT DestVT = TLI->getRegisterType(NewVT); 905 RegisterVT = DestVT; 906 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 907 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 908 909 // Otherwise, promotion or legal types use the same number of registers as 910 // the vector decimated to the appropriate level. 911 return NumVectorRegs; 912 } 913 914 /// isLegalRC - Return true if the value types that can be represented by the 915 /// specified register class are all legal. 916 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 917 const TargetRegisterClass &RC) const { 918 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 919 if (isTypeLegal(*I)) 920 return true; 921 return false; 922 } 923 924 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 925 /// sequence of memory operands that is recognized by PrologEpilogInserter. 926 MachineBasicBlock * 927 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 928 MachineBasicBlock *MBB) const { 929 MachineInstr *MI = &InitialMI; 930 MachineFunction &MF = *MI->getMF(); 931 MachineFrameInfo &MFI = MF.getFrameInfo(); 932 933 // We're handling multiple types of operands here: 934 // PATCHPOINT MetaArgs - live-in, read only, direct 935 // STATEPOINT Deopt Spill - live-through, read only, indirect 936 // STATEPOINT Deopt Alloca - live-through, read only, direct 937 // (We're currently conservative and mark the deopt slots read/write in 938 // practice.) 939 // STATEPOINT GC Spill - live-through, read/write, indirect 940 // STATEPOINT GC Alloca - live-through, read/write, direct 941 // The live-in vs live-through is handled already (the live through ones are 942 // all stack slots), but we need to handle the different type of stackmap 943 // operands and memory effects here. 944 945 // MI changes inside this loop as we grow operands. 946 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 947 MachineOperand &MO = MI->getOperand(OperIdx); 948 if (!MO.isFI()) 949 continue; 950 951 // foldMemoryOperand builds a new MI after replacing a single FI operand 952 // with the canonical set of five x86 addressing-mode operands. 953 int FI = MO.getIndex(); 954 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 955 956 // Copy operands before the frame-index. 957 for (unsigned i = 0; i < OperIdx; ++i) 958 MIB.add(MI->getOperand(i)); 959 // Add frame index operands recognized by stackmaps.cpp 960 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 961 // indirect-mem-ref tag, size, #FI, offset. 962 // Used for spills inserted by StatepointLowering. This codepath is not 963 // used for patchpoints/stackmaps at all, for these spilling is done via 964 // foldMemoryOperand callback only. 965 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 966 MIB.addImm(StackMaps::IndirectMemRefOp); 967 MIB.addImm(MFI.getObjectSize(FI)); 968 MIB.add(MI->getOperand(OperIdx)); 969 MIB.addImm(0); 970 } else { 971 // direct-mem-ref tag, #FI, offset. 972 // Used by patchpoint, and direct alloca arguments to statepoints 973 MIB.addImm(StackMaps::DirectMemRefOp); 974 MIB.add(MI->getOperand(OperIdx)); 975 MIB.addImm(0); 976 } 977 // Copy the operands after the frame index. 978 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 979 MIB.add(MI->getOperand(i)); 980 981 // Inherit previous memory operands. 982 MIB.cloneMemRefs(*MI); 983 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 984 985 // Add a new memory operand for this FI. 986 assert(MFI.getObjectOffset(FI) != -1); 987 988 auto Flags = MachineMemOperand::MOLoad; 989 if (MI->getOpcode() == TargetOpcode::STATEPOINT) { 990 Flags |= MachineMemOperand::MOStore; 991 Flags |= MachineMemOperand::MOVolatile; 992 } 993 MachineMemOperand *MMO = MF.getMachineMemOperand( 994 MachinePointerInfo::getFixedStack(MF, FI), Flags, 995 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 996 MIB->addMemOperand(MF, MMO); 997 998 // Replace the instruction and update the operand index. 999 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1000 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1001 MI->eraseFromParent(); 1002 MI = MIB; 1003 } 1004 return MBB; 1005 } 1006 1007 MachineBasicBlock * 1008 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1009 MachineBasicBlock *MBB) const { 1010 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1011 "Called emitXRayCustomEvent on the wrong MI!"); 1012 auto &MF = *MI.getMF(); 1013 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1014 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1015 MIB.add(MI.getOperand(OpIdx)); 1016 1017 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1018 MI.eraseFromParent(); 1019 return MBB; 1020 } 1021 1022 MachineBasicBlock * 1023 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1024 MachineBasicBlock *MBB) const { 1025 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1026 "Called emitXRayTypedEvent on the wrong MI!"); 1027 auto &MF = *MI.getMF(); 1028 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1029 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1030 MIB.add(MI.getOperand(OpIdx)); 1031 1032 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1033 MI.eraseFromParent(); 1034 return MBB; 1035 } 1036 1037 /// findRepresentativeClass - Return the largest legal super-reg register class 1038 /// of the register class for the specified type and its associated "cost". 1039 // This function is in TargetLowering because it uses RegClassForVT which would 1040 // need to be moved to TargetRegisterInfo and would necessitate moving 1041 // isTypeLegal over as well - a massive change that would just require 1042 // TargetLowering having a TargetRegisterInfo class member that it would use. 1043 std::pair<const TargetRegisterClass *, uint8_t> 1044 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1045 MVT VT) const { 1046 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1047 if (!RC) 1048 return std::make_pair(RC, 0); 1049 1050 // Compute the set of all super-register classes. 1051 BitVector SuperRegRC(TRI->getNumRegClasses()); 1052 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1053 SuperRegRC.setBitsInMask(RCI.getMask()); 1054 1055 // Find the first legal register class with the largest spill size. 1056 const TargetRegisterClass *BestRC = RC; 1057 for (unsigned i : SuperRegRC.set_bits()) { 1058 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1059 // We want the largest possible spill size. 1060 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1061 continue; 1062 if (!isLegalRC(*TRI, *SuperRC)) 1063 continue; 1064 BestRC = SuperRC; 1065 } 1066 return std::make_pair(BestRC, 1); 1067 } 1068 1069 /// computeRegisterProperties - Once all of the register classes are added, 1070 /// this allows us to compute derived properties we expose. 1071 void TargetLoweringBase::computeRegisterProperties( 1072 const TargetRegisterInfo *TRI) { 1073 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1074 "Too many value types for ValueTypeActions to hold!"); 1075 1076 // Everything defaults to needing one register. 1077 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1078 NumRegistersForVT[i] = 1; 1079 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1080 } 1081 // ...except isVoid, which doesn't need any registers. 1082 NumRegistersForVT[MVT::isVoid] = 0; 1083 1084 // Find the largest integer register class. 1085 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1086 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1087 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1088 1089 // Every integer value type larger than this largest register takes twice as 1090 // many registers to represent as the previous ValueType. 1091 for (unsigned ExpandedReg = LargestIntReg + 1; 1092 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1093 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1094 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1095 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1096 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1097 TypeExpandInteger); 1098 } 1099 1100 // Inspect all of the ValueType's smaller than the largest integer 1101 // register to see which ones need promotion. 1102 unsigned LegalIntReg = LargestIntReg; 1103 for (unsigned IntReg = LargestIntReg - 1; 1104 IntReg >= (unsigned)MVT::i1; --IntReg) { 1105 MVT IVT = (MVT::SimpleValueType)IntReg; 1106 if (isTypeLegal(IVT)) { 1107 LegalIntReg = IntReg; 1108 } else { 1109 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1110 (MVT::SimpleValueType)LegalIntReg; 1111 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1112 } 1113 } 1114 1115 // ppcf128 type is really two f64's. 1116 if (!isTypeLegal(MVT::ppcf128)) { 1117 if (isTypeLegal(MVT::f64)) { 1118 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1119 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1120 TransformToType[MVT::ppcf128] = MVT::f64; 1121 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1122 } else { 1123 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1124 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1125 TransformToType[MVT::ppcf128] = MVT::i128; 1126 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1127 } 1128 } 1129 1130 // Decide how to handle f128. If the target does not have native f128 support, 1131 // expand it to i128 and we will be generating soft float library calls. 1132 if (!isTypeLegal(MVT::f128)) { 1133 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1134 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1135 TransformToType[MVT::f128] = MVT::i128; 1136 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1137 } 1138 1139 // Decide how to handle f64. If the target does not have native f64 support, 1140 // expand it to i64 and we will be generating soft float library calls. 1141 if (!isTypeLegal(MVT::f64)) { 1142 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1143 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1144 TransformToType[MVT::f64] = MVT::i64; 1145 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1146 } 1147 1148 // Decide how to handle f32. If the target does not have native f32 support, 1149 // expand it to i32 and we will be generating soft float library calls. 1150 if (!isTypeLegal(MVT::f32)) { 1151 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1152 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1153 TransformToType[MVT::f32] = MVT::i32; 1154 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1155 } 1156 1157 // Decide how to handle f16. If the target does not have native f16 support, 1158 // promote it to f32, because there are no f16 library calls (except for 1159 // conversions). 1160 if (!isTypeLegal(MVT::f16)) { 1161 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1162 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1163 TransformToType[MVT::f16] = MVT::f32; 1164 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1165 } 1166 1167 // Loop over all of the vector value types to see which need transformations. 1168 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1169 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1170 MVT VT = (MVT::SimpleValueType) i; 1171 if (isTypeLegal(VT)) 1172 continue; 1173 1174 MVT EltVT = VT.getVectorElementType(); 1175 unsigned NElts = VT.getVectorNumElements(); 1176 bool IsLegalWiderType = false; 1177 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1178 switch (PreferredAction) { 1179 case TypePromoteInteger: 1180 // Try to promote the elements of integer vectors. If no legal 1181 // promotion was found, fall through to the widen-vector method. 1182 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1183 MVT SVT = (MVT::SimpleValueType) nVT; 1184 // Promote vectors of integers to vectors with the same number 1185 // of elements, with a wider element type. 1186 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1187 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1188 TransformToType[i] = SVT; 1189 RegisterTypeForVT[i] = SVT; 1190 NumRegistersForVT[i] = 1; 1191 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1192 IsLegalWiderType = true; 1193 break; 1194 } 1195 } 1196 if (IsLegalWiderType) 1197 break; 1198 LLVM_FALLTHROUGH; 1199 1200 case TypeWidenVector: 1201 // Try to widen the vector. 1202 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1203 MVT SVT = (MVT::SimpleValueType) nVT; 1204 if (SVT.getVectorElementType() == EltVT 1205 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1206 TransformToType[i] = SVT; 1207 RegisterTypeForVT[i] = SVT; 1208 NumRegistersForVT[i] = 1; 1209 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1210 IsLegalWiderType = true; 1211 break; 1212 } 1213 } 1214 if (IsLegalWiderType) 1215 break; 1216 LLVM_FALLTHROUGH; 1217 1218 case TypeSplitVector: 1219 case TypeScalarizeVector: { 1220 MVT IntermediateVT; 1221 MVT RegisterVT; 1222 unsigned NumIntermediates; 1223 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1224 NumIntermediates, RegisterVT, this); 1225 RegisterTypeForVT[i] = RegisterVT; 1226 1227 MVT NVT = VT.getPow2VectorType(); 1228 if (NVT == VT) { 1229 // Type is already a power of 2. The default action is to split. 1230 TransformToType[i] = MVT::Other; 1231 if (PreferredAction == TypeScalarizeVector) 1232 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1233 else if (PreferredAction == TypeSplitVector) 1234 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1235 else 1236 // Set type action according to the number of elements. 1237 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1238 : TypeSplitVector); 1239 } else { 1240 TransformToType[i] = NVT; 1241 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1242 } 1243 break; 1244 } 1245 default: 1246 llvm_unreachable("Unknown vector legalization action!"); 1247 } 1248 } 1249 1250 // Determine the 'representative' register class for each value type. 1251 // An representative register class is the largest (meaning one which is 1252 // not a sub-register class / subreg register class) legal register class for 1253 // a group of value types. For example, on i386, i8, i16, and i32 1254 // representative would be GR32; while on x86_64 it's GR64. 1255 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1256 const TargetRegisterClass* RRC; 1257 uint8_t Cost; 1258 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1259 RepRegClassForVT[i] = RRC; 1260 RepRegClassCostForVT[i] = Cost; 1261 } 1262 } 1263 1264 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1265 EVT VT) const { 1266 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1267 return getPointerTy(DL).SimpleTy; 1268 } 1269 1270 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1271 return MVT::i32; // return the default value 1272 } 1273 1274 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1275 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1276 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1277 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1278 /// 1279 /// This method returns the number of registers needed, and the VT for each 1280 /// register. It also returns the VT and quantity of the intermediate values 1281 /// before they are promoted/expanded. 1282 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1283 EVT &IntermediateVT, 1284 unsigned &NumIntermediates, 1285 MVT &RegisterVT) const { 1286 unsigned NumElts = VT.getVectorNumElements(); 1287 1288 // If there is a wider vector type with the same element type as this one, 1289 // or a promoted vector type that has the same number of elements which 1290 // are wider, then we should convert to that legal vector type. 1291 // This handles things like <2 x float> -> <4 x float> and 1292 // <4 x i1> -> <4 x i32>. 1293 LegalizeTypeAction TA = getTypeAction(Context, VT); 1294 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1295 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1296 if (isTypeLegal(RegisterEVT)) { 1297 IntermediateVT = RegisterEVT; 1298 RegisterVT = RegisterEVT.getSimpleVT(); 1299 NumIntermediates = 1; 1300 return 1; 1301 } 1302 } 1303 1304 // Figure out the right, legal destination reg to copy into. 1305 EVT EltTy = VT.getVectorElementType(); 1306 1307 unsigned NumVectorRegs = 1; 1308 1309 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1310 // could break down into LHS/RHS like LegalizeDAG does. 1311 if (!isPowerOf2_32(NumElts)) { 1312 NumVectorRegs = NumElts; 1313 NumElts = 1; 1314 } 1315 1316 // Divide the input until we get to a supported size. This will always 1317 // end with a scalar if the target doesn't support vectors. 1318 while (NumElts > 1 && !isTypeLegal( 1319 EVT::getVectorVT(Context, EltTy, NumElts))) { 1320 NumElts >>= 1; 1321 NumVectorRegs <<= 1; 1322 } 1323 1324 NumIntermediates = NumVectorRegs; 1325 1326 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1327 if (!isTypeLegal(NewVT)) 1328 NewVT = EltTy; 1329 IntermediateVT = NewVT; 1330 1331 MVT DestVT = getRegisterType(Context, NewVT); 1332 RegisterVT = DestVT; 1333 unsigned NewVTSize = NewVT.getSizeInBits(); 1334 1335 // Convert sizes such as i33 to i64. 1336 if (!isPowerOf2_32(NewVTSize)) 1337 NewVTSize = NextPowerOf2(NewVTSize); 1338 1339 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1340 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1341 1342 // Otherwise, promotion or legal types use the same number of registers as 1343 // the vector decimated to the appropriate level. 1344 return NumVectorRegs; 1345 } 1346 1347 /// Get the EVTs and ArgFlags collections that represent the legalized return 1348 /// type of the given function. This does not require a DAG or a return value, 1349 /// and is suitable for use before any DAGs for the function are constructed. 1350 /// TODO: Move this out of TargetLowering.cpp. 1351 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1352 AttributeList attr, 1353 SmallVectorImpl<ISD::OutputArg> &Outs, 1354 const TargetLowering &TLI, const DataLayout &DL) { 1355 SmallVector<EVT, 4> ValueVTs; 1356 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1357 unsigned NumValues = ValueVTs.size(); 1358 if (NumValues == 0) return; 1359 1360 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1361 EVT VT = ValueVTs[j]; 1362 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1363 1364 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1365 ExtendKind = ISD::SIGN_EXTEND; 1366 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1367 ExtendKind = ISD::ZERO_EXTEND; 1368 1369 // FIXME: C calling convention requires the return type to be promoted to 1370 // at least 32-bit. But this is not necessary for non-C calling 1371 // conventions. The frontend should mark functions whose return values 1372 // require promoting with signext or zeroext attributes. 1373 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1374 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1375 if (VT.bitsLT(MinVT)) 1376 VT = MinVT; 1377 } 1378 1379 unsigned NumParts = 1380 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1381 MVT PartVT = 1382 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1383 1384 // 'inreg' on function refers to return value 1385 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1386 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1387 Flags.setInReg(); 1388 1389 // Propagate extension type if any 1390 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1391 Flags.setSExt(); 1392 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1393 Flags.setZExt(); 1394 1395 for (unsigned i = 0; i < NumParts; ++i) 1396 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0)); 1397 } 1398 } 1399 1400 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1401 /// function arguments in the caller parameter area. This is the actual 1402 /// alignment, not its logarithm. 1403 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1404 const DataLayout &DL) const { 1405 return DL.getABITypeAlignment(Ty); 1406 } 1407 1408 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1409 const DataLayout &DL, EVT VT, 1410 unsigned AddrSpace, 1411 unsigned Alignment, 1412 bool *Fast) const { 1413 // Check if the specified alignment is sufficient based on the data layout. 1414 // TODO: While using the data layout works in practice, a better solution 1415 // would be to implement this check directly (make this a virtual function). 1416 // For example, the ABI alignment may change based on software platform while 1417 // this function should only be affected by hardware implementation. 1418 Type *Ty = VT.getTypeForEVT(Context); 1419 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1420 // Assume that an access that meets the ABI-specified alignment is fast. 1421 if (Fast != nullptr) 1422 *Fast = true; 1423 return true; 1424 } 1425 1426 // This is a misaligned access. 1427 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast); 1428 } 1429 1430 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1431 return BranchProbability(MinPercentageForPredictableBranch, 100); 1432 } 1433 1434 //===----------------------------------------------------------------------===// 1435 // TargetTransformInfo Helpers 1436 //===----------------------------------------------------------------------===// 1437 1438 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1439 enum InstructionOpcodes { 1440 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1441 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1442 #include "llvm/IR/Instruction.def" 1443 }; 1444 switch (static_cast<InstructionOpcodes>(Opcode)) { 1445 case Ret: return 0; 1446 case Br: return 0; 1447 case Switch: return 0; 1448 case IndirectBr: return 0; 1449 case Invoke: return 0; 1450 case Resume: return 0; 1451 case Unreachable: return 0; 1452 case CleanupRet: return 0; 1453 case CatchRet: return 0; 1454 case CatchPad: return 0; 1455 case CatchSwitch: return 0; 1456 case CleanupPad: return 0; 1457 case FNeg: return ISD::FNEG; 1458 case Add: return ISD::ADD; 1459 case FAdd: return ISD::FADD; 1460 case Sub: return ISD::SUB; 1461 case FSub: return ISD::FSUB; 1462 case Mul: return ISD::MUL; 1463 case FMul: return ISD::FMUL; 1464 case UDiv: return ISD::UDIV; 1465 case SDiv: return ISD::SDIV; 1466 case FDiv: return ISD::FDIV; 1467 case URem: return ISD::UREM; 1468 case SRem: return ISD::SREM; 1469 case FRem: return ISD::FREM; 1470 case Shl: return ISD::SHL; 1471 case LShr: return ISD::SRL; 1472 case AShr: return ISD::SRA; 1473 case And: return ISD::AND; 1474 case Or: return ISD::OR; 1475 case Xor: return ISD::XOR; 1476 case Alloca: return 0; 1477 case Load: return ISD::LOAD; 1478 case Store: return ISD::STORE; 1479 case GetElementPtr: return 0; 1480 case Fence: return 0; 1481 case AtomicCmpXchg: return 0; 1482 case AtomicRMW: return 0; 1483 case Trunc: return ISD::TRUNCATE; 1484 case ZExt: return ISD::ZERO_EXTEND; 1485 case SExt: return ISD::SIGN_EXTEND; 1486 case FPToUI: return ISD::FP_TO_UINT; 1487 case FPToSI: return ISD::FP_TO_SINT; 1488 case UIToFP: return ISD::UINT_TO_FP; 1489 case SIToFP: return ISD::SINT_TO_FP; 1490 case FPTrunc: return ISD::FP_ROUND; 1491 case FPExt: return ISD::FP_EXTEND; 1492 case PtrToInt: return ISD::BITCAST; 1493 case IntToPtr: return ISD::BITCAST; 1494 case BitCast: return ISD::BITCAST; 1495 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1496 case ICmp: return ISD::SETCC; 1497 case FCmp: return ISD::SETCC; 1498 case PHI: return 0; 1499 case Call: return 0; 1500 case Select: return ISD::SELECT; 1501 case UserOp1: return 0; 1502 case UserOp2: return 0; 1503 case VAArg: return 0; 1504 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1505 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1506 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1507 case ExtractValue: return ISD::MERGE_VALUES; 1508 case InsertValue: return ISD::MERGE_VALUES; 1509 case LandingPad: return 0; 1510 } 1511 1512 llvm_unreachable("Unknown instruction type encountered!"); 1513 } 1514 1515 std::pair<int, MVT> 1516 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1517 Type *Ty) const { 1518 LLVMContext &C = Ty->getContext(); 1519 EVT MTy = getValueType(DL, Ty); 1520 1521 int Cost = 1; 1522 // We keep legalizing the type until we find a legal kind. We assume that 1523 // the only operation that costs anything is the split. After splitting 1524 // we need to handle two types. 1525 while (true) { 1526 LegalizeKind LK = getTypeConversion(C, MTy); 1527 1528 if (LK.first == TypeLegal) 1529 return std::make_pair(Cost, MTy.getSimpleVT()); 1530 1531 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1532 Cost *= 2; 1533 1534 // Do not loop with f128 type. 1535 if (MTy == LK.second) 1536 return std::make_pair(Cost, MTy.getSimpleVT()); 1537 1538 // Keep legalizing the type. 1539 MTy = LK.second; 1540 } 1541 } 1542 1543 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1544 bool UseTLS) const { 1545 // compiler-rt provides a variable with a magic name. Targets that do not 1546 // link with compiler-rt may also provide such a variable. 1547 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1548 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1549 auto UnsafeStackPtr = 1550 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1551 1552 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1553 1554 if (!UnsafeStackPtr) { 1555 auto TLSModel = UseTLS ? 1556 GlobalValue::InitialExecTLSModel : 1557 GlobalValue::NotThreadLocal; 1558 // The global variable is not defined yet, define it ourselves. 1559 // We use the initial-exec TLS model because we do not support the 1560 // variable living anywhere other than in the main executable. 1561 UnsafeStackPtr = new GlobalVariable( 1562 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1563 UnsafeStackPtrVar, nullptr, TLSModel); 1564 } else { 1565 // The variable exists, check its type and attributes. 1566 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1567 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1568 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1569 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1570 (UseTLS ? "" : "not ") + "be thread-local"); 1571 } 1572 return UnsafeStackPtr; 1573 } 1574 1575 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1576 if (!TM.getTargetTriple().isAndroid()) 1577 return getDefaultSafeStackPointerLocation(IRB, true); 1578 1579 // Android provides a libc function to retrieve the address of the current 1580 // thread's unsafe stack pointer. 1581 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1582 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1583 Value *Fn = M->getOrInsertFunction("__safestack_pointer_address", 1584 StackPtrTy->getPointerTo(0)); 1585 return IRB.CreateCall(Fn); 1586 } 1587 1588 //===----------------------------------------------------------------------===// 1589 // Loop Strength Reduction hooks 1590 //===----------------------------------------------------------------------===// 1591 1592 /// isLegalAddressingMode - Return true if the addressing mode represented 1593 /// by AM is legal for this target, for a load/store of the specified type. 1594 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1595 const AddrMode &AM, Type *Ty, 1596 unsigned AS, Instruction *I) const { 1597 // The default implementation of this implements a conservative RISCy, r+r and 1598 // r+i addr mode. 1599 1600 // Allows a sign-extended 16-bit immediate field. 1601 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1602 return false; 1603 1604 // No global is ever allowed as a base. 1605 if (AM.BaseGV) 1606 return false; 1607 1608 // Only support r+r, 1609 switch (AM.Scale) { 1610 case 0: // "r+i" or just "i", depending on HasBaseReg. 1611 break; 1612 case 1: 1613 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1614 return false; 1615 // Otherwise we have r+r or r+i. 1616 break; 1617 case 2: 1618 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1619 return false; 1620 // Allow 2*r as r+r. 1621 break; 1622 default: // Don't allow n * r 1623 return false; 1624 } 1625 1626 return true; 1627 } 1628 1629 //===----------------------------------------------------------------------===// 1630 // Stack Protector 1631 //===----------------------------------------------------------------------===// 1632 1633 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1634 // so that SelectionDAG handle SSP. 1635 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1636 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1637 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1638 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1639 return M.getOrInsertGlobal("__guard_local", PtrTy); 1640 } 1641 return nullptr; 1642 } 1643 1644 // Currently only support "standard" __stack_chk_guard. 1645 // TODO: add LOAD_STACK_GUARD support. 1646 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1647 if (!M.getNamedValue("__stack_chk_guard")) 1648 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1649 GlobalVariable::ExternalLinkage, 1650 nullptr, "__stack_chk_guard"); 1651 } 1652 1653 // Currently only support "standard" __stack_chk_guard. 1654 // TODO: add LOAD_STACK_GUARD support. 1655 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1656 return M.getNamedValue("__stack_chk_guard"); 1657 } 1658 1659 Value *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1660 return nullptr; 1661 } 1662 1663 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1664 return MinimumJumpTableEntries; 1665 } 1666 1667 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1668 MinimumJumpTableEntries = Val; 1669 } 1670 1671 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1672 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1673 } 1674 1675 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1676 return MaximumJumpTableSize; 1677 } 1678 1679 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1680 MaximumJumpTableSize = Val; 1681 } 1682 1683 //===----------------------------------------------------------------------===// 1684 // Reciprocal Estimates 1685 //===----------------------------------------------------------------------===// 1686 1687 /// Get the reciprocal estimate attribute string for a function that will 1688 /// override the target defaults. 1689 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1690 const Function &F = MF.getFunction(); 1691 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1692 } 1693 1694 /// Construct a string for the given reciprocal operation of the given type. 1695 /// This string should match the corresponding option to the front-end's 1696 /// "-mrecip" flag assuming those strings have been passed through in an 1697 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1698 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1699 std::string Name = VT.isVector() ? "vec-" : ""; 1700 1701 Name += IsSqrt ? "sqrt" : "div"; 1702 1703 // TODO: Handle "half" or other float types? 1704 if (VT.getScalarType() == MVT::f64) { 1705 Name += "d"; 1706 } else { 1707 assert(VT.getScalarType() == MVT::f32 && 1708 "Unexpected FP type for reciprocal estimate"); 1709 Name += "f"; 1710 } 1711 1712 return Name; 1713 } 1714 1715 /// Return the character position and value (a single numeric character) of a 1716 /// customized refinement operation in the input string if it exists. Return 1717 /// false if there is no customized refinement step count. 1718 static bool parseRefinementStep(StringRef In, size_t &Position, 1719 uint8_t &Value) { 1720 const char RefStepToken = ':'; 1721 Position = In.find(RefStepToken); 1722 if (Position == StringRef::npos) 1723 return false; 1724 1725 StringRef RefStepString = In.substr(Position + 1); 1726 // Allow exactly one numeric character for the additional refinement 1727 // step parameter. 1728 if (RefStepString.size() == 1) { 1729 char RefStepChar = RefStepString[0]; 1730 if (RefStepChar >= '0' && RefStepChar <= '9') { 1731 Value = RefStepChar - '0'; 1732 return true; 1733 } 1734 } 1735 report_fatal_error("Invalid refinement step for -recip."); 1736 } 1737 1738 /// For the input attribute string, return one of the ReciprocalEstimate enum 1739 /// status values (enabled, disabled, or not specified) for this operation on 1740 /// the specified data type. 1741 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1742 if (Override.empty()) 1743 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1744 1745 SmallVector<StringRef, 4> OverrideVector; 1746 Override.split(OverrideVector, ','); 1747 unsigned NumArgs = OverrideVector.size(); 1748 1749 // Check if "all", "none", or "default" was specified. 1750 if (NumArgs == 1) { 1751 // Look for an optional setting of the number of refinement steps needed 1752 // for this type of reciprocal operation. 1753 size_t RefPos; 1754 uint8_t RefSteps; 1755 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1756 // Split the string for further processing. 1757 Override = Override.substr(0, RefPos); 1758 } 1759 1760 // All reciprocal types are enabled. 1761 if (Override == "all") 1762 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1763 1764 // All reciprocal types are disabled. 1765 if (Override == "none") 1766 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1767 1768 // Target defaults for enablement are used. 1769 if (Override == "default") 1770 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1771 } 1772 1773 // The attribute string may omit the size suffix ('f'/'d'). 1774 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1775 std::string VTNameNoSize = VTName; 1776 VTNameNoSize.pop_back(); 1777 static const char DisabledPrefix = '!'; 1778 1779 for (StringRef RecipType : OverrideVector) { 1780 size_t RefPos; 1781 uint8_t RefSteps; 1782 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1783 RecipType = RecipType.substr(0, RefPos); 1784 1785 // Ignore the disablement token for string matching. 1786 bool IsDisabled = RecipType[0] == DisabledPrefix; 1787 if (IsDisabled) 1788 RecipType = RecipType.substr(1); 1789 1790 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1791 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1792 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1793 } 1794 1795 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1796 } 1797 1798 /// For the input attribute string, return the customized refinement step count 1799 /// for this operation on the specified data type. If the step count does not 1800 /// exist, return the ReciprocalEstimate enum value for unspecified. 1801 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1802 if (Override.empty()) 1803 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1804 1805 SmallVector<StringRef, 4> OverrideVector; 1806 Override.split(OverrideVector, ','); 1807 unsigned NumArgs = OverrideVector.size(); 1808 1809 // Check if "all", "default", or "none" was specified. 1810 if (NumArgs == 1) { 1811 // Look for an optional setting of the number of refinement steps needed 1812 // for this type of reciprocal operation. 1813 size_t RefPos; 1814 uint8_t RefSteps; 1815 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1816 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1817 1818 // Split the string for further processing. 1819 Override = Override.substr(0, RefPos); 1820 assert(Override != "none" && 1821 "Disabled reciprocals, but specifed refinement steps?"); 1822 1823 // If this is a general override, return the specified number of steps. 1824 if (Override == "all" || Override == "default") 1825 return RefSteps; 1826 } 1827 1828 // The attribute string may omit the size suffix ('f'/'d'). 1829 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1830 std::string VTNameNoSize = VTName; 1831 VTNameNoSize.pop_back(); 1832 1833 for (StringRef RecipType : OverrideVector) { 1834 size_t RefPos; 1835 uint8_t RefSteps; 1836 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1837 continue; 1838 1839 RecipType = RecipType.substr(0, RefPos); 1840 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1841 return RefSteps; 1842 } 1843 1844 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1845 } 1846 1847 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1848 MachineFunction &MF) const { 1849 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1850 } 1851 1852 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1853 MachineFunction &MF) const { 1854 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1855 } 1856 1857 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1858 MachineFunction &MF) const { 1859 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1860 } 1861 1862 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1863 MachineFunction &MF) const { 1864 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1865 } 1866 1867 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1868 MF.getRegInfo().freezeReservedRegs(MF); 1869 } 1870