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::FMINNAN, VT, Expand); 606 setOperationAction(ISD::FMAXNAN, 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::SADDSAT, VT, Expand); 614 615 // Overflow operations default to expand 616 setOperationAction(ISD::SADDO, VT, Expand); 617 setOperationAction(ISD::SSUBO, VT, Expand); 618 setOperationAction(ISD::UADDO, VT, Expand); 619 setOperationAction(ISD::USUBO, VT, Expand); 620 setOperationAction(ISD::SMULO, VT, Expand); 621 setOperationAction(ISD::UMULO, VT, Expand); 622 623 // ADDCARRY operations default to expand 624 setOperationAction(ISD::ADDCARRY, VT, Expand); 625 setOperationAction(ISD::SUBCARRY, VT, Expand); 626 setOperationAction(ISD::SETCCCARRY, VT, Expand); 627 628 // ADDC/ADDE/SUBC/SUBE default to expand. 629 setOperationAction(ISD::ADDC, VT, Expand); 630 setOperationAction(ISD::ADDE, VT, Expand); 631 setOperationAction(ISD::SUBC, VT, Expand); 632 setOperationAction(ISD::SUBE, VT, Expand); 633 634 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 635 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 636 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 637 638 setOperationAction(ISD::BITREVERSE, VT, Expand); 639 640 // These library functions default to expand. 641 setOperationAction(ISD::FROUND, VT, Expand); 642 setOperationAction(ISD::FPOWI, VT, Expand); 643 644 // These operations default to expand for vector types. 645 if (VT.isVector()) { 646 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 647 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 648 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 649 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 650 } 651 652 // For most targets @llvm.get.dynamic.area.offset just returns 0. 653 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 654 } 655 656 // Most targets ignore the @llvm.prefetch intrinsic. 657 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 658 659 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 660 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 661 662 // ConstantFP nodes default to expand. Targets can either change this to 663 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 664 // to optimize expansions for certain constants. 665 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 666 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 667 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 668 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 669 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 670 671 // These library functions default to expand. 672 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 673 setOperationAction(ISD::FCBRT, VT, Expand); 674 setOperationAction(ISD::FLOG , VT, Expand); 675 setOperationAction(ISD::FLOG2, VT, Expand); 676 setOperationAction(ISD::FLOG10, VT, Expand); 677 setOperationAction(ISD::FEXP , VT, Expand); 678 setOperationAction(ISD::FEXP2, VT, Expand); 679 setOperationAction(ISD::FFLOOR, VT, Expand); 680 setOperationAction(ISD::FNEARBYINT, VT, Expand); 681 setOperationAction(ISD::FCEIL, VT, Expand); 682 setOperationAction(ISD::FRINT, VT, Expand); 683 setOperationAction(ISD::FTRUNC, VT, Expand); 684 setOperationAction(ISD::FROUND, VT, Expand); 685 } 686 687 // Default ISD::TRAP to expand (which turns it into abort). 688 setOperationAction(ISD::TRAP, MVT::Other, Expand); 689 690 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 691 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 692 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 693 } 694 695 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 696 EVT) const { 697 return MVT::getIntegerVT(8 * DL.getPointerSize(0)); 698 } 699 700 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 701 bool LegalTypes) const { 702 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 703 if (LHSTy.isVector()) 704 return LHSTy; 705 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 706 : getPointerTy(DL); 707 } 708 709 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 710 assert(isTypeLegal(VT)); 711 switch (Op) { 712 default: 713 return false; 714 case ISD::SDIV: 715 case ISD::UDIV: 716 case ISD::SREM: 717 case ISD::UREM: 718 return true; 719 } 720 } 721 722 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 723 // If the command-line option was specified, ignore this request. 724 if (!JumpIsExpensiveOverride.getNumOccurrences()) 725 JumpIsExpensive = isExpensive; 726 } 727 728 TargetLoweringBase::LegalizeKind 729 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 730 // If this is a simple type, use the ComputeRegisterProp mechanism. 731 if (VT.isSimple()) { 732 MVT SVT = VT.getSimpleVT(); 733 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 734 MVT NVT = TransformToType[SVT.SimpleTy]; 735 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 736 737 assert((LA == TypeLegal || LA == TypeSoftenFloat || 738 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) && 739 "Promote may not follow Expand or Promote"); 740 741 if (LA == TypeSplitVector) 742 return LegalizeKind(LA, 743 EVT::getVectorVT(Context, SVT.getVectorElementType(), 744 SVT.getVectorNumElements() / 2)); 745 if (LA == TypeScalarizeVector) 746 return LegalizeKind(LA, SVT.getVectorElementType()); 747 return LegalizeKind(LA, NVT); 748 } 749 750 // Handle Extended Scalar Types. 751 if (!VT.isVector()) { 752 assert(VT.isInteger() && "Float types must be simple"); 753 unsigned BitSize = VT.getSizeInBits(); 754 // First promote to a power-of-two size, then expand if necessary. 755 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 756 EVT NVT = VT.getRoundIntegerType(Context); 757 assert(NVT != VT && "Unable to round integer VT"); 758 LegalizeKind NextStep = getTypeConversion(Context, NVT); 759 // Avoid multi-step promotion. 760 if (NextStep.first == TypePromoteInteger) 761 return NextStep; 762 // Return rounded integer type. 763 return LegalizeKind(TypePromoteInteger, NVT); 764 } 765 766 return LegalizeKind(TypeExpandInteger, 767 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 768 } 769 770 // Handle vector types. 771 unsigned NumElts = VT.getVectorNumElements(); 772 EVT EltVT = VT.getVectorElementType(); 773 774 // Vectors with only one element are always scalarized. 775 if (NumElts == 1) 776 return LegalizeKind(TypeScalarizeVector, EltVT); 777 778 // Try to widen vector elements until the element type is a power of two and 779 // promote it to a legal type later on, for example: 780 // <3 x i8> -> <4 x i8> -> <4 x i32> 781 if (EltVT.isInteger()) { 782 // Vectors with a number of elements that is not a power of two are always 783 // widened, for example <3 x i8> -> <4 x i8>. 784 if (!VT.isPow2VectorType()) { 785 NumElts = (unsigned)NextPowerOf2(NumElts); 786 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 787 return LegalizeKind(TypeWidenVector, NVT); 788 } 789 790 // Examine the element type. 791 LegalizeKind LK = getTypeConversion(Context, EltVT); 792 793 // If type is to be expanded, split the vector. 794 // <4 x i140> -> <2 x i140> 795 if (LK.first == TypeExpandInteger) 796 return LegalizeKind(TypeSplitVector, 797 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 798 799 // Promote the integer element types until a legal vector type is found 800 // or until the element integer type is too big. If a legal type was not 801 // found, fallback to the usual mechanism of widening/splitting the 802 // vector. 803 EVT OldEltVT = EltVT; 804 while (true) { 805 // Increase the bitwidth of the element to the next pow-of-two 806 // (which is greater than 8 bits). 807 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 808 .getRoundIntegerType(Context); 809 810 // Stop trying when getting a non-simple element type. 811 // Note that vector elements may be greater than legal vector element 812 // types. Example: X86 XMM registers hold 64bit element on 32bit 813 // systems. 814 if (!EltVT.isSimple()) 815 break; 816 817 // Build a new vector type and check if it is legal. 818 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 819 // Found a legal promoted vector type. 820 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 821 return LegalizeKind(TypePromoteInteger, 822 EVT::getVectorVT(Context, EltVT, NumElts)); 823 } 824 825 // Reset the type to the unexpanded type if we did not find a legal vector 826 // type with a promoted vector element type. 827 EltVT = OldEltVT; 828 } 829 830 // Try to widen the vector until a legal type is found. 831 // If there is no wider legal type, split the vector. 832 while (true) { 833 // Round up to the next power of 2. 834 NumElts = (unsigned)NextPowerOf2(NumElts); 835 836 // If there is no simple vector type with this many elements then there 837 // cannot be a larger legal vector type. Note that this assumes that 838 // there are no skipped intermediate vector types in the simple types. 839 if (!EltVT.isSimple()) 840 break; 841 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 842 if (LargerVector == MVT()) 843 break; 844 845 // If this type is legal then widen the vector. 846 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 847 return LegalizeKind(TypeWidenVector, LargerVector); 848 } 849 850 // Widen odd vectors to next power of two. 851 if (!VT.isPow2VectorType()) { 852 EVT NVT = VT.getPow2VectorType(Context); 853 return LegalizeKind(TypeWidenVector, NVT); 854 } 855 856 // Vectors with illegal element types are expanded. 857 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 858 return LegalizeKind(TypeSplitVector, NVT); 859 } 860 861 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 862 unsigned &NumIntermediates, 863 MVT &RegisterVT, 864 TargetLoweringBase *TLI) { 865 // Figure out the right, legal destination reg to copy into. 866 unsigned NumElts = VT.getVectorNumElements(); 867 MVT EltTy = VT.getVectorElementType(); 868 869 unsigned NumVectorRegs = 1; 870 871 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 872 // could break down into LHS/RHS like LegalizeDAG does. 873 if (!isPowerOf2_32(NumElts)) { 874 NumVectorRegs = NumElts; 875 NumElts = 1; 876 } 877 878 // Divide the input until we get to a supported size. This will always 879 // end with a scalar if the target doesn't support vectors. 880 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 881 NumElts >>= 1; 882 NumVectorRegs <<= 1; 883 } 884 885 NumIntermediates = NumVectorRegs; 886 887 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 888 if (!TLI->isTypeLegal(NewVT)) 889 NewVT = EltTy; 890 IntermediateVT = NewVT; 891 892 unsigned NewVTSize = NewVT.getSizeInBits(); 893 894 // Convert sizes such as i33 to i64. 895 if (!isPowerOf2_32(NewVTSize)) 896 NewVTSize = NextPowerOf2(NewVTSize); 897 898 MVT DestVT = TLI->getRegisterType(NewVT); 899 RegisterVT = DestVT; 900 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 901 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 902 903 // Otherwise, promotion or legal types use the same number of registers as 904 // the vector decimated to the appropriate level. 905 return NumVectorRegs; 906 } 907 908 /// isLegalRC - Return true if the value types that can be represented by the 909 /// specified register class are all legal. 910 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 911 const TargetRegisterClass &RC) const { 912 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 913 if (isTypeLegal(*I)) 914 return true; 915 return false; 916 } 917 918 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 919 /// sequence of memory operands that is recognized by PrologEpilogInserter. 920 MachineBasicBlock * 921 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 922 MachineBasicBlock *MBB) const { 923 MachineInstr *MI = &InitialMI; 924 MachineFunction &MF = *MI->getMF(); 925 MachineFrameInfo &MFI = MF.getFrameInfo(); 926 927 // We're handling multiple types of operands here: 928 // PATCHPOINT MetaArgs - live-in, read only, direct 929 // STATEPOINT Deopt Spill - live-through, read only, indirect 930 // STATEPOINT Deopt Alloca - live-through, read only, direct 931 // (We're currently conservative and mark the deopt slots read/write in 932 // practice.) 933 // STATEPOINT GC Spill - live-through, read/write, indirect 934 // STATEPOINT GC Alloca - live-through, read/write, direct 935 // The live-in vs live-through is handled already (the live through ones are 936 // all stack slots), but we need to handle the different type of stackmap 937 // operands and memory effects here. 938 939 // MI changes inside this loop as we grow operands. 940 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 941 MachineOperand &MO = MI->getOperand(OperIdx); 942 if (!MO.isFI()) 943 continue; 944 945 // foldMemoryOperand builds a new MI after replacing a single FI operand 946 // with the canonical set of five x86 addressing-mode operands. 947 int FI = MO.getIndex(); 948 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 949 950 // Copy operands before the frame-index. 951 for (unsigned i = 0; i < OperIdx; ++i) 952 MIB.add(MI->getOperand(i)); 953 // Add frame index operands recognized by stackmaps.cpp 954 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 955 // indirect-mem-ref tag, size, #FI, offset. 956 // Used for spills inserted by StatepointLowering. This codepath is not 957 // used for patchpoints/stackmaps at all, for these spilling is done via 958 // foldMemoryOperand callback only. 959 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 960 MIB.addImm(StackMaps::IndirectMemRefOp); 961 MIB.addImm(MFI.getObjectSize(FI)); 962 MIB.add(MI->getOperand(OperIdx)); 963 MIB.addImm(0); 964 } else { 965 // direct-mem-ref tag, #FI, offset. 966 // Used by patchpoint, and direct alloca arguments to statepoints 967 MIB.addImm(StackMaps::DirectMemRefOp); 968 MIB.add(MI->getOperand(OperIdx)); 969 MIB.addImm(0); 970 } 971 // Copy the operands after the frame index. 972 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 973 MIB.add(MI->getOperand(i)); 974 975 // Inherit previous memory operands. 976 MIB.cloneMemRefs(*MI); 977 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 978 979 // Add a new memory operand for this FI. 980 assert(MFI.getObjectOffset(FI) != -1); 981 982 auto Flags = MachineMemOperand::MOLoad; 983 if (MI->getOpcode() == TargetOpcode::STATEPOINT) { 984 Flags |= MachineMemOperand::MOStore; 985 Flags |= MachineMemOperand::MOVolatile; 986 } 987 MachineMemOperand *MMO = MF.getMachineMemOperand( 988 MachinePointerInfo::getFixedStack(MF, FI), Flags, 989 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 990 MIB->addMemOperand(MF, MMO); 991 992 // Replace the instruction and update the operand index. 993 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 994 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 995 MI->eraseFromParent(); 996 MI = MIB; 997 } 998 return MBB; 999 } 1000 1001 MachineBasicBlock * 1002 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1003 MachineBasicBlock *MBB) const { 1004 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1005 "Called emitXRayCustomEvent on the wrong MI!"); 1006 auto &MF = *MI.getMF(); 1007 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1008 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1009 MIB.add(MI.getOperand(OpIdx)); 1010 1011 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1012 MI.eraseFromParent(); 1013 return MBB; 1014 } 1015 1016 MachineBasicBlock * 1017 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1018 MachineBasicBlock *MBB) const { 1019 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1020 "Called emitXRayTypedEvent on the wrong MI!"); 1021 auto &MF = *MI.getMF(); 1022 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1023 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1024 MIB.add(MI.getOperand(OpIdx)); 1025 1026 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1027 MI.eraseFromParent(); 1028 return MBB; 1029 } 1030 1031 /// findRepresentativeClass - Return the largest legal super-reg register class 1032 /// of the register class for the specified type and its associated "cost". 1033 // This function is in TargetLowering because it uses RegClassForVT which would 1034 // need to be moved to TargetRegisterInfo and would necessitate moving 1035 // isTypeLegal over as well - a massive change that would just require 1036 // TargetLowering having a TargetRegisterInfo class member that it would use. 1037 std::pair<const TargetRegisterClass *, uint8_t> 1038 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1039 MVT VT) const { 1040 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1041 if (!RC) 1042 return std::make_pair(RC, 0); 1043 1044 // Compute the set of all super-register classes. 1045 BitVector SuperRegRC(TRI->getNumRegClasses()); 1046 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1047 SuperRegRC.setBitsInMask(RCI.getMask()); 1048 1049 // Find the first legal register class with the largest spill size. 1050 const TargetRegisterClass *BestRC = RC; 1051 for (unsigned i : SuperRegRC.set_bits()) { 1052 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1053 // We want the largest possible spill size. 1054 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1055 continue; 1056 if (!isLegalRC(*TRI, *SuperRC)) 1057 continue; 1058 BestRC = SuperRC; 1059 } 1060 return std::make_pair(BestRC, 1); 1061 } 1062 1063 /// computeRegisterProperties - Once all of the register classes are added, 1064 /// this allows us to compute derived properties we expose. 1065 void TargetLoweringBase::computeRegisterProperties( 1066 const TargetRegisterInfo *TRI) { 1067 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1068 "Too many value types for ValueTypeActions to hold!"); 1069 1070 // Everything defaults to needing one register. 1071 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1072 NumRegistersForVT[i] = 1; 1073 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1074 } 1075 // ...except isVoid, which doesn't need any registers. 1076 NumRegistersForVT[MVT::isVoid] = 0; 1077 1078 // Find the largest integer register class. 1079 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1080 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1081 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1082 1083 // Every integer value type larger than this largest register takes twice as 1084 // many registers to represent as the previous ValueType. 1085 for (unsigned ExpandedReg = LargestIntReg + 1; 1086 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1087 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1088 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1089 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1090 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1091 TypeExpandInteger); 1092 } 1093 1094 // Inspect all of the ValueType's smaller than the largest integer 1095 // register to see which ones need promotion. 1096 unsigned LegalIntReg = LargestIntReg; 1097 for (unsigned IntReg = LargestIntReg - 1; 1098 IntReg >= (unsigned)MVT::i1; --IntReg) { 1099 MVT IVT = (MVT::SimpleValueType)IntReg; 1100 if (isTypeLegal(IVT)) { 1101 LegalIntReg = IntReg; 1102 } else { 1103 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1104 (const MVT::SimpleValueType)LegalIntReg; 1105 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1106 } 1107 } 1108 1109 // ppcf128 type is really two f64's. 1110 if (!isTypeLegal(MVT::ppcf128)) { 1111 if (isTypeLegal(MVT::f64)) { 1112 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1113 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1114 TransformToType[MVT::ppcf128] = MVT::f64; 1115 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1116 } else { 1117 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1118 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1119 TransformToType[MVT::ppcf128] = MVT::i128; 1120 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1121 } 1122 } 1123 1124 // Decide how to handle f128. If the target does not have native f128 support, 1125 // expand it to i128 and we will be generating soft float library calls. 1126 if (!isTypeLegal(MVT::f128)) { 1127 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1128 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1129 TransformToType[MVT::f128] = MVT::i128; 1130 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1131 } 1132 1133 // Decide how to handle f64. If the target does not have native f64 support, 1134 // expand it to i64 and we will be generating soft float library calls. 1135 if (!isTypeLegal(MVT::f64)) { 1136 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1137 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1138 TransformToType[MVT::f64] = MVT::i64; 1139 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1140 } 1141 1142 // Decide how to handle f32. If the target does not have native f32 support, 1143 // expand it to i32 and we will be generating soft float library calls. 1144 if (!isTypeLegal(MVT::f32)) { 1145 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1146 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1147 TransformToType[MVT::f32] = MVT::i32; 1148 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1149 } 1150 1151 // Decide how to handle f16. If the target does not have native f16 support, 1152 // promote it to f32, because there are no f16 library calls (except for 1153 // conversions). 1154 if (!isTypeLegal(MVT::f16)) { 1155 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1156 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1157 TransformToType[MVT::f16] = MVT::f32; 1158 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1159 } 1160 1161 // Loop over all of the vector value types to see which need transformations. 1162 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1163 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1164 MVT VT = (MVT::SimpleValueType) i; 1165 if (isTypeLegal(VT)) 1166 continue; 1167 1168 MVT EltVT = VT.getVectorElementType(); 1169 unsigned NElts = VT.getVectorNumElements(); 1170 bool IsLegalWiderType = false; 1171 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1172 switch (PreferredAction) { 1173 case TypePromoteInteger: 1174 // Try to promote the elements of integer vectors. If no legal 1175 // promotion was found, fall through to the widen-vector method. 1176 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1177 MVT SVT = (MVT::SimpleValueType) nVT; 1178 // Promote vectors of integers to vectors with the same number 1179 // of elements, with a wider element type. 1180 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1181 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1182 TransformToType[i] = SVT; 1183 RegisterTypeForVT[i] = SVT; 1184 NumRegistersForVT[i] = 1; 1185 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1186 IsLegalWiderType = true; 1187 break; 1188 } 1189 } 1190 if (IsLegalWiderType) 1191 break; 1192 LLVM_FALLTHROUGH; 1193 1194 case TypeWidenVector: 1195 // Try to widen the vector. 1196 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1197 MVT SVT = (MVT::SimpleValueType) nVT; 1198 if (SVT.getVectorElementType() == EltVT 1199 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1200 TransformToType[i] = SVT; 1201 RegisterTypeForVT[i] = SVT; 1202 NumRegistersForVT[i] = 1; 1203 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1204 IsLegalWiderType = true; 1205 break; 1206 } 1207 } 1208 if (IsLegalWiderType) 1209 break; 1210 LLVM_FALLTHROUGH; 1211 1212 case TypeSplitVector: 1213 case TypeScalarizeVector: { 1214 MVT IntermediateVT; 1215 MVT RegisterVT; 1216 unsigned NumIntermediates; 1217 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1218 NumIntermediates, RegisterVT, this); 1219 RegisterTypeForVT[i] = RegisterVT; 1220 1221 MVT NVT = VT.getPow2VectorType(); 1222 if (NVT == VT) { 1223 // Type is already a power of 2. The default action is to split. 1224 TransformToType[i] = MVT::Other; 1225 if (PreferredAction == TypeScalarizeVector) 1226 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1227 else if (PreferredAction == TypeSplitVector) 1228 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1229 else 1230 // Set type action according to the number of elements. 1231 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1232 : TypeSplitVector); 1233 } else { 1234 TransformToType[i] = NVT; 1235 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1236 } 1237 break; 1238 } 1239 default: 1240 llvm_unreachable("Unknown vector legalization action!"); 1241 } 1242 } 1243 1244 // Determine the 'representative' register class for each value type. 1245 // An representative register class is the largest (meaning one which is 1246 // not a sub-register class / subreg register class) legal register class for 1247 // a group of value types. For example, on i386, i8, i16, and i32 1248 // representative would be GR32; while on x86_64 it's GR64. 1249 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1250 const TargetRegisterClass* RRC; 1251 uint8_t Cost; 1252 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1253 RepRegClassForVT[i] = RRC; 1254 RepRegClassCostForVT[i] = Cost; 1255 } 1256 } 1257 1258 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1259 EVT VT) const { 1260 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1261 return getPointerTy(DL).SimpleTy; 1262 } 1263 1264 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1265 return MVT::i32; // return the default value 1266 } 1267 1268 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1269 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1270 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1271 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1272 /// 1273 /// This method returns the number of registers needed, and the VT for each 1274 /// register. It also returns the VT and quantity of the intermediate values 1275 /// before they are promoted/expanded. 1276 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1277 EVT &IntermediateVT, 1278 unsigned &NumIntermediates, 1279 MVT &RegisterVT) const { 1280 unsigned NumElts = VT.getVectorNumElements(); 1281 1282 // If there is a wider vector type with the same element type as this one, 1283 // or a promoted vector type that has the same number of elements which 1284 // are wider, then we should convert to that legal vector type. 1285 // This handles things like <2 x float> -> <4 x float> and 1286 // <4 x i1> -> <4 x i32>. 1287 LegalizeTypeAction TA = getTypeAction(Context, VT); 1288 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1289 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1290 if (isTypeLegal(RegisterEVT)) { 1291 IntermediateVT = RegisterEVT; 1292 RegisterVT = RegisterEVT.getSimpleVT(); 1293 NumIntermediates = 1; 1294 return 1; 1295 } 1296 } 1297 1298 // Figure out the right, legal destination reg to copy into. 1299 EVT EltTy = VT.getVectorElementType(); 1300 1301 unsigned NumVectorRegs = 1; 1302 1303 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1304 // could break down into LHS/RHS like LegalizeDAG does. 1305 if (!isPowerOf2_32(NumElts)) { 1306 NumVectorRegs = NumElts; 1307 NumElts = 1; 1308 } 1309 1310 // Divide the input until we get to a supported size. This will always 1311 // end with a scalar if the target doesn't support vectors. 1312 while (NumElts > 1 && !isTypeLegal( 1313 EVT::getVectorVT(Context, EltTy, NumElts))) { 1314 NumElts >>= 1; 1315 NumVectorRegs <<= 1; 1316 } 1317 1318 NumIntermediates = NumVectorRegs; 1319 1320 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1321 if (!isTypeLegal(NewVT)) 1322 NewVT = EltTy; 1323 IntermediateVT = NewVT; 1324 1325 MVT DestVT = getRegisterType(Context, NewVT); 1326 RegisterVT = DestVT; 1327 unsigned NewVTSize = NewVT.getSizeInBits(); 1328 1329 // Convert sizes such as i33 to i64. 1330 if (!isPowerOf2_32(NewVTSize)) 1331 NewVTSize = NextPowerOf2(NewVTSize); 1332 1333 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1334 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1335 1336 // Otherwise, promotion or legal types use the same number of registers as 1337 // the vector decimated to the appropriate level. 1338 return NumVectorRegs; 1339 } 1340 1341 /// Get the EVTs and ArgFlags collections that represent the legalized return 1342 /// type of the given function. This does not require a DAG or a return value, 1343 /// and is suitable for use before any DAGs for the function are constructed. 1344 /// TODO: Move this out of TargetLowering.cpp. 1345 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1346 AttributeList attr, 1347 SmallVectorImpl<ISD::OutputArg> &Outs, 1348 const TargetLowering &TLI, const DataLayout &DL) { 1349 SmallVector<EVT, 4> ValueVTs; 1350 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1351 unsigned NumValues = ValueVTs.size(); 1352 if (NumValues == 0) return; 1353 1354 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1355 EVT VT = ValueVTs[j]; 1356 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1357 1358 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1359 ExtendKind = ISD::SIGN_EXTEND; 1360 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1361 ExtendKind = ISD::ZERO_EXTEND; 1362 1363 // FIXME: C calling convention requires the return type to be promoted to 1364 // at least 32-bit. But this is not necessary for non-C calling 1365 // conventions. The frontend should mark functions whose return values 1366 // require promoting with signext or zeroext attributes. 1367 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1368 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1369 if (VT.bitsLT(MinVT)) 1370 VT = MinVT; 1371 } 1372 1373 unsigned NumParts = 1374 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1375 MVT PartVT = 1376 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1377 1378 // 'inreg' on function refers to return value 1379 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1380 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1381 Flags.setInReg(); 1382 1383 // Propagate extension type if any 1384 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1385 Flags.setSExt(); 1386 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1387 Flags.setZExt(); 1388 1389 for (unsigned i = 0; i < NumParts; ++i) 1390 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0)); 1391 } 1392 } 1393 1394 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1395 /// function arguments in the caller parameter area. This is the actual 1396 /// alignment, not its logarithm. 1397 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1398 const DataLayout &DL) const { 1399 return DL.getABITypeAlignment(Ty); 1400 } 1401 1402 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1403 const DataLayout &DL, EVT VT, 1404 unsigned AddrSpace, 1405 unsigned Alignment, 1406 bool *Fast) const { 1407 // Check if the specified alignment is sufficient based on the data layout. 1408 // TODO: While using the data layout works in practice, a better solution 1409 // would be to implement this check directly (make this a virtual function). 1410 // For example, the ABI alignment may change based on software platform while 1411 // this function should only be affected by hardware implementation. 1412 Type *Ty = VT.getTypeForEVT(Context); 1413 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1414 // Assume that an access that meets the ABI-specified alignment is fast. 1415 if (Fast != nullptr) 1416 *Fast = true; 1417 return true; 1418 } 1419 1420 // This is a misaligned access. 1421 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast); 1422 } 1423 1424 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1425 return BranchProbability(MinPercentageForPredictableBranch, 100); 1426 } 1427 1428 //===----------------------------------------------------------------------===// 1429 // TargetTransformInfo Helpers 1430 //===----------------------------------------------------------------------===// 1431 1432 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1433 enum InstructionOpcodes { 1434 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1435 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1436 #include "llvm/IR/Instruction.def" 1437 }; 1438 switch (static_cast<InstructionOpcodes>(Opcode)) { 1439 case Ret: return 0; 1440 case Br: return 0; 1441 case Switch: return 0; 1442 case IndirectBr: return 0; 1443 case Invoke: return 0; 1444 case Resume: return 0; 1445 case Unreachable: return 0; 1446 case CleanupRet: return 0; 1447 case CatchRet: return 0; 1448 case CatchPad: return 0; 1449 case CatchSwitch: return 0; 1450 case CleanupPad: return 0; 1451 case Add: return ISD::ADD; 1452 case FAdd: return ISD::FADD; 1453 case Sub: return ISD::SUB; 1454 case FSub: return ISD::FSUB; 1455 case Mul: return ISD::MUL; 1456 case FMul: return ISD::FMUL; 1457 case UDiv: return ISD::UDIV; 1458 case SDiv: return ISD::SDIV; 1459 case FDiv: return ISD::FDIV; 1460 case URem: return ISD::UREM; 1461 case SRem: return ISD::SREM; 1462 case FRem: return ISD::FREM; 1463 case Shl: return ISD::SHL; 1464 case LShr: return ISD::SRL; 1465 case AShr: return ISD::SRA; 1466 case And: return ISD::AND; 1467 case Or: return ISD::OR; 1468 case Xor: return ISD::XOR; 1469 case Alloca: return 0; 1470 case Load: return ISD::LOAD; 1471 case Store: return ISD::STORE; 1472 case GetElementPtr: return 0; 1473 case Fence: return 0; 1474 case AtomicCmpXchg: return 0; 1475 case AtomicRMW: return 0; 1476 case Trunc: return ISD::TRUNCATE; 1477 case ZExt: return ISD::ZERO_EXTEND; 1478 case SExt: return ISD::SIGN_EXTEND; 1479 case FPToUI: return ISD::FP_TO_UINT; 1480 case FPToSI: return ISD::FP_TO_SINT; 1481 case UIToFP: return ISD::UINT_TO_FP; 1482 case SIToFP: return ISD::SINT_TO_FP; 1483 case FPTrunc: return ISD::FP_ROUND; 1484 case FPExt: return ISD::FP_EXTEND; 1485 case PtrToInt: return ISD::BITCAST; 1486 case IntToPtr: return ISD::BITCAST; 1487 case BitCast: return ISD::BITCAST; 1488 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1489 case ICmp: return ISD::SETCC; 1490 case FCmp: return ISD::SETCC; 1491 case PHI: return 0; 1492 case Call: return 0; 1493 case Select: return ISD::SELECT; 1494 case UserOp1: return 0; 1495 case UserOp2: return 0; 1496 case VAArg: return 0; 1497 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1498 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1499 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1500 case ExtractValue: return ISD::MERGE_VALUES; 1501 case InsertValue: return ISD::MERGE_VALUES; 1502 case LandingPad: return 0; 1503 } 1504 1505 llvm_unreachable("Unknown instruction type encountered!"); 1506 } 1507 1508 std::pair<int, MVT> 1509 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1510 Type *Ty) const { 1511 LLVMContext &C = Ty->getContext(); 1512 EVT MTy = getValueType(DL, Ty); 1513 1514 int Cost = 1; 1515 // We keep legalizing the type until we find a legal kind. We assume that 1516 // the only operation that costs anything is the split. After splitting 1517 // we need to handle two types. 1518 while (true) { 1519 LegalizeKind LK = getTypeConversion(C, MTy); 1520 1521 if (LK.first == TypeLegal) 1522 return std::make_pair(Cost, MTy.getSimpleVT()); 1523 1524 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1525 Cost *= 2; 1526 1527 // Do not loop with f128 type. 1528 if (MTy == LK.second) 1529 return std::make_pair(Cost, MTy.getSimpleVT()); 1530 1531 // Keep legalizing the type. 1532 MTy = LK.second; 1533 } 1534 } 1535 1536 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1537 bool UseTLS) const { 1538 // compiler-rt provides a variable with a magic name. Targets that do not 1539 // link with compiler-rt may also provide such a variable. 1540 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1541 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1542 auto UnsafeStackPtr = 1543 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1544 1545 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1546 1547 if (!UnsafeStackPtr) { 1548 auto TLSModel = UseTLS ? 1549 GlobalValue::InitialExecTLSModel : 1550 GlobalValue::NotThreadLocal; 1551 // The global variable is not defined yet, define it ourselves. 1552 // We use the initial-exec TLS model because we do not support the 1553 // variable living anywhere other than in the main executable. 1554 UnsafeStackPtr = new GlobalVariable( 1555 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1556 UnsafeStackPtrVar, nullptr, TLSModel); 1557 } else { 1558 // The variable exists, check its type and attributes. 1559 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1560 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1561 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1562 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1563 (UseTLS ? "" : "not ") + "be thread-local"); 1564 } 1565 return UnsafeStackPtr; 1566 } 1567 1568 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1569 if (!TM.getTargetTriple().isAndroid()) 1570 return getDefaultSafeStackPointerLocation(IRB, true); 1571 1572 // Android provides a libc function to retrieve the address of the current 1573 // thread's unsafe stack pointer. 1574 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1575 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1576 Value *Fn = M->getOrInsertFunction("__safestack_pointer_address", 1577 StackPtrTy->getPointerTo(0)); 1578 return IRB.CreateCall(Fn); 1579 } 1580 1581 //===----------------------------------------------------------------------===// 1582 // Loop Strength Reduction hooks 1583 //===----------------------------------------------------------------------===// 1584 1585 /// isLegalAddressingMode - Return true if the addressing mode represented 1586 /// by AM is legal for this target, for a load/store of the specified type. 1587 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1588 const AddrMode &AM, Type *Ty, 1589 unsigned AS, Instruction *I) const { 1590 // The default implementation of this implements a conservative RISCy, r+r and 1591 // r+i addr mode. 1592 1593 // Allows a sign-extended 16-bit immediate field. 1594 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1595 return false; 1596 1597 // No global is ever allowed as a base. 1598 if (AM.BaseGV) 1599 return false; 1600 1601 // Only support r+r, 1602 switch (AM.Scale) { 1603 case 0: // "r+i" or just "i", depending on HasBaseReg. 1604 break; 1605 case 1: 1606 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1607 return false; 1608 // Otherwise we have r+r or r+i. 1609 break; 1610 case 2: 1611 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1612 return false; 1613 // Allow 2*r as r+r. 1614 break; 1615 default: // Don't allow n * r 1616 return false; 1617 } 1618 1619 return true; 1620 } 1621 1622 //===----------------------------------------------------------------------===// 1623 // Stack Protector 1624 //===----------------------------------------------------------------------===// 1625 1626 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1627 // so that SelectionDAG handle SSP. 1628 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1629 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1630 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1631 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1632 return M.getOrInsertGlobal("__guard_local", PtrTy); 1633 } 1634 return nullptr; 1635 } 1636 1637 // Currently only support "standard" __stack_chk_guard. 1638 // TODO: add LOAD_STACK_GUARD support. 1639 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1640 if (!M.getNamedValue("__stack_chk_guard")) 1641 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1642 GlobalVariable::ExternalLinkage, 1643 nullptr, "__stack_chk_guard"); 1644 } 1645 1646 // Currently only support "standard" __stack_chk_guard. 1647 // TODO: add LOAD_STACK_GUARD support. 1648 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1649 return M.getNamedValue("__stack_chk_guard"); 1650 } 1651 1652 Value *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1653 return nullptr; 1654 } 1655 1656 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1657 return MinimumJumpTableEntries; 1658 } 1659 1660 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1661 MinimumJumpTableEntries = Val; 1662 } 1663 1664 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1665 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1666 } 1667 1668 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1669 return MaximumJumpTableSize; 1670 } 1671 1672 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1673 MaximumJumpTableSize = Val; 1674 } 1675 1676 //===----------------------------------------------------------------------===// 1677 // Reciprocal Estimates 1678 //===----------------------------------------------------------------------===// 1679 1680 /// Get the reciprocal estimate attribute string for a function that will 1681 /// override the target defaults. 1682 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1683 const Function &F = MF.getFunction(); 1684 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1685 } 1686 1687 /// Construct a string for the given reciprocal operation of the given type. 1688 /// This string should match the corresponding option to the front-end's 1689 /// "-mrecip" flag assuming those strings have been passed through in an 1690 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1691 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1692 std::string Name = VT.isVector() ? "vec-" : ""; 1693 1694 Name += IsSqrt ? "sqrt" : "div"; 1695 1696 // TODO: Handle "half" or other float types? 1697 if (VT.getScalarType() == MVT::f64) { 1698 Name += "d"; 1699 } else { 1700 assert(VT.getScalarType() == MVT::f32 && 1701 "Unexpected FP type for reciprocal estimate"); 1702 Name += "f"; 1703 } 1704 1705 return Name; 1706 } 1707 1708 /// Return the character position and value (a single numeric character) of a 1709 /// customized refinement operation in the input string if it exists. Return 1710 /// false if there is no customized refinement step count. 1711 static bool parseRefinementStep(StringRef In, size_t &Position, 1712 uint8_t &Value) { 1713 const char RefStepToken = ':'; 1714 Position = In.find(RefStepToken); 1715 if (Position == StringRef::npos) 1716 return false; 1717 1718 StringRef RefStepString = In.substr(Position + 1); 1719 // Allow exactly one numeric character for the additional refinement 1720 // step parameter. 1721 if (RefStepString.size() == 1) { 1722 char RefStepChar = RefStepString[0]; 1723 if (RefStepChar >= '0' && RefStepChar <= '9') { 1724 Value = RefStepChar - '0'; 1725 return true; 1726 } 1727 } 1728 report_fatal_error("Invalid refinement step for -recip."); 1729 } 1730 1731 /// For the input attribute string, return one of the ReciprocalEstimate enum 1732 /// status values (enabled, disabled, or not specified) for this operation on 1733 /// the specified data type. 1734 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1735 if (Override.empty()) 1736 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1737 1738 SmallVector<StringRef, 4> OverrideVector; 1739 Override.split(OverrideVector, ','); 1740 unsigned NumArgs = OverrideVector.size(); 1741 1742 // Check if "all", "none", or "default" was specified. 1743 if (NumArgs == 1) { 1744 // Look for an optional setting of the number of refinement steps needed 1745 // for this type of reciprocal operation. 1746 size_t RefPos; 1747 uint8_t RefSteps; 1748 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1749 // Split the string for further processing. 1750 Override = Override.substr(0, RefPos); 1751 } 1752 1753 // All reciprocal types are enabled. 1754 if (Override == "all") 1755 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1756 1757 // All reciprocal types are disabled. 1758 if (Override == "none") 1759 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1760 1761 // Target defaults for enablement are used. 1762 if (Override == "default") 1763 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1764 } 1765 1766 // The attribute string may omit the size suffix ('f'/'d'). 1767 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1768 std::string VTNameNoSize = VTName; 1769 VTNameNoSize.pop_back(); 1770 static const char DisabledPrefix = '!'; 1771 1772 for (StringRef RecipType : OverrideVector) { 1773 size_t RefPos; 1774 uint8_t RefSteps; 1775 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1776 RecipType = RecipType.substr(0, RefPos); 1777 1778 // Ignore the disablement token for string matching. 1779 bool IsDisabled = RecipType[0] == DisabledPrefix; 1780 if (IsDisabled) 1781 RecipType = RecipType.substr(1); 1782 1783 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1784 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1785 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1786 } 1787 1788 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1789 } 1790 1791 /// For the input attribute string, return the customized refinement step count 1792 /// for this operation on the specified data type. If the step count does not 1793 /// exist, return the ReciprocalEstimate enum value for unspecified. 1794 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1795 if (Override.empty()) 1796 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1797 1798 SmallVector<StringRef, 4> OverrideVector; 1799 Override.split(OverrideVector, ','); 1800 unsigned NumArgs = OverrideVector.size(); 1801 1802 // Check if "all", "default", or "none" was specified. 1803 if (NumArgs == 1) { 1804 // Look for an optional setting of the number of refinement steps needed 1805 // for this type of reciprocal operation. 1806 size_t RefPos; 1807 uint8_t RefSteps; 1808 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1809 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1810 1811 // Split the string for further processing. 1812 Override = Override.substr(0, RefPos); 1813 assert(Override != "none" && 1814 "Disabled reciprocals, but specifed refinement steps?"); 1815 1816 // If this is a general override, return the specified number of steps. 1817 if (Override == "all" || Override == "default") 1818 return RefSteps; 1819 } 1820 1821 // The attribute string may omit the size suffix ('f'/'d'). 1822 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1823 std::string VTNameNoSize = VTName; 1824 VTNameNoSize.pop_back(); 1825 1826 for (StringRef RecipType : OverrideVector) { 1827 size_t RefPos; 1828 uint8_t RefSteps; 1829 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1830 continue; 1831 1832 RecipType = RecipType.substr(0, RefPos); 1833 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1834 return RefSteps; 1835 } 1836 1837 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1838 } 1839 1840 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1841 MachineFunction &MF) const { 1842 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1843 } 1844 1845 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1846 MachineFunction &MF) const { 1847 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1848 } 1849 1850 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1851 MachineFunction &MF) const { 1852 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1853 } 1854 1855 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1856 MachineFunction &MF) const { 1857 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1858 } 1859 1860 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1861 MF.getRegInfo().freezeReservedRegs(MF); 1862 } 1863