1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This implements the TargetLoweringBase class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/ADT/BitVector.h" 15 #include "llvm/ADT/STLExtras.h" 16 #include "llvm/ADT/SmallVector.h" 17 #include "llvm/ADT/StringExtras.h" 18 #include "llvm/ADT/StringRef.h" 19 #include "llvm/ADT/Triple.h" 20 #include "llvm/ADT/Twine.h" 21 #include "llvm/CodeGen/Analysis.h" 22 #include "llvm/CodeGen/ISDOpcodes.h" 23 #include "llvm/CodeGen/MachineBasicBlock.h" 24 #include "llvm/CodeGen/MachineFrameInfo.h" 25 #include "llvm/CodeGen/MachineFunction.h" 26 #include "llvm/CodeGen/MachineInstr.h" 27 #include "llvm/CodeGen/MachineInstrBuilder.h" 28 #include "llvm/CodeGen/MachineMemOperand.h" 29 #include "llvm/CodeGen/MachineOperand.h" 30 #include "llvm/CodeGen/MachineRegisterInfo.h" 31 #include "llvm/CodeGen/MachineValueType.h" 32 #include "llvm/CodeGen/RuntimeLibcalls.h" 33 #include "llvm/CodeGen/StackMaps.h" 34 #include "llvm/CodeGen/TargetLowering.h" 35 #include "llvm/CodeGen/TargetOpcodes.h" 36 #include "llvm/CodeGen/TargetRegisterInfo.h" 37 #include "llvm/CodeGen/ValueTypes.h" 38 #include "llvm/IR/Attributes.h" 39 #include "llvm/IR/CallingConv.h" 40 #include "llvm/IR/DataLayout.h" 41 #include "llvm/IR/DerivedTypes.h" 42 #include "llvm/IR/Function.h" 43 #include "llvm/IR/GlobalValue.h" 44 #include "llvm/IR/GlobalVariable.h" 45 #include "llvm/IR/IRBuilder.h" 46 #include "llvm/IR/Module.h" 47 #include "llvm/IR/Type.h" 48 #include "llvm/Support/BranchProbability.h" 49 #include "llvm/Support/Casting.h" 50 #include "llvm/Support/CommandLine.h" 51 #include "llvm/Support/Compiler.h" 52 #include "llvm/Support/ErrorHandling.h" 53 #include "llvm/Support/MathExtras.h" 54 #include "llvm/Target/TargetMachine.h" 55 #include <algorithm> 56 #include <cassert> 57 #include <cstddef> 58 #include <cstdint> 59 #include <cstring> 60 #include <iterator> 61 #include <string> 62 #include <tuple> 63 #include <utility> 64 65 using namespace llvm; 66 67 static cl::opt<bool> JumpIsExpensiveOverride( 68 "jump-is-expensive", cl::init(false), 69 cl::desc("Do not create extra branches to split comparison logic."), 70 cl::Hidden); 71 72 static cl::opt<unsigned> MinimumJumpTableEntries 73 ("min-jump-table-entries", cl::init(4), cl::Hidden, 74 cl::desc("Set minimum number of entries to use a jump table.")); 75 76 static cl::opt<unsigned> MaximumJumpTableSize 77 ("max-jump-table-size", cl::init(0), cl::Hidden, 78 cl::desc("Set maximum size of jump tables; zero for no limit.")); 79 80 /// Minimum jump table density for normal functions. 81 static cl::opt<unsigned> 82 JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden, 83 cl::desc("Minimum density for building a jump table in " 84 "a normal function")); 85 86 /// Minimum jump table density for -Os or -Oz functions. 87 static cl::opt<unsigned> OptsizeJumpTableDensity( 88 "optsize-jump-table-density", cl::init(40), cl::Hidden, 89 cl::desc("Minimum density for building a jump table in " 90 "an optsize function")); 91 92 static bool darwinHasSinCos(const Triple &TT) { 93 assert(TT.isOSDarwin() && "should be called with darwin triple"); 94 // Don't bother with 32 bit x86. 95 if (TT.getArch() == Triple::x86) 96 return false; 97 // Macos < 10.9 has no sincos_stret. 98 if (TT.isMacOSX()) 99 return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit(); 100 // iOS < 7.0 has no sincos_stret. 101 if (TT.isiOS()) 102 return !TT.isOSVersionLT(7, 0); 103 // Any other darwin such as WatchOS/TvOS is new enough. 104 return true; 105 } 106 107 // Although this default value is arbitrary, it is not random. It is assumed 108 // that a condition that evaluates the same way by a higher percentage than this 109 // is best represented as control flow. Therefore, the default value N should be 110 // set such that the win from N% correct executions is greater than the loss 111 // from (100 - N)% mispredicted executions for the majority of intended targets. 112 static cl::opt<int> MinPercentageForPredictableBranch( 113 "min-predictable-branch", cl::init(99), 114 cl::desc("Minimum percentage (0-100) that a condition must be either true " 115 "or false to assume that the condition is predictable"), 116 cl::Hidden); 117 118 void TargetLoweringBase::InitLibcalls(const Triple &TT) { 119 #define HANDLE_LIBCALL(code, name) \ 120 setLibcallName(RTLIB::code, name); 121 #include "llvm/CodeGen/RuntimeLibcalls.def" 122 #undef HANDLE_LIBCALL 123 // Initialize calling conventions to their default. 124 for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC) 125 setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C); 126 127 // A few names are different on particular architectures or environments. 128 if (TT.isOSDarwin()) { 129 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead 130 // of the gnueabi-style __gnu_*_ieee. 131 // FIXME: What about other targets? 132 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2"); 133 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2"); 134 135 // Some darwins have an optimized __bzero/bzero function. 136 switch (TT.getArch()) { 137 case Triple::x86: 138 case Triple::x86_64: 139 if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6)) 140 setLibcallName(RTLIB::BZERO, "__bzero"); 141 break; 142 case Triple::aarch64: 143 setLibcallName(RTLIB::BZERO, "bzero"); 144 break; 145 default: 146 break; 147 } 148 149 if (darwinHasSinCos(TT)) { 150 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret"); 151 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret"); 152 if (TT.isWatchABI()) { 153 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32, 154 CallingConv::ARM_AAPCS_VFP); 155 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64, 156 CallingConv::ARM_AAPCS_VFP); 157 } 158 } 159 } else { 160 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee"); 161 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee"); 162 } 163 164 if (TT.isGNUEnvironment() || TT.isOSFuchsia()) { 165 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 166 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 167 setLibcallName(RTLIB::SINCOS_F80, "sincosl"); 168 setLibcallName(RTLIB::SINCOS_F128, "sincosl"); 169 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl"); 170 } 171 172 if (TT.isOSOpenBSD()) { 173 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr); 174 } 175 } 176 177 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 178 /// UNKNOWN_LIBCALL if there is none. 179 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 180 if (OpVT == MVT::f16) { 181 if (RetVT == MVT::f32) 182 return FPEXT_F16_F32; 183 } else if (OpVT == MVT::f32) { 184 if (RetVT == MVT::f64) 185 return FPEXT_F32_F64; 186 if (RetVT == MVT::f128) 187 return FPEXT_F32_F128; 188 if (RetVT == MVT::ppcf128) 189 return FPEXT_F32_PPCF128; 190 } else if (OpVT == MVT::f64) { 191 if (RetVT == MVT::f128) 192 return FPEXT_F64_F128; 193 else if (RetVT == MVT::ppcf128) 194 return FPEXT_F64_PPCF128; 195 } else if (OpVT == MVT::f80) { 196 if (RetVT == MVT::f128) 197 return FPEXT_F80_F128; 198 } 199 200 return UNKNOWN_LIBCALL; 201 } 202 203 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 204 /// UNKNOWN_LIBCALL if there is none. 205 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 206 if (RetVT == MVT::f16) { 207 if (OpVT == MVT::f32) 208 return FPROUND_F32_F16; 209 if (OpVT == MVT::f64) 210 return FPROUND_F64_F16; 211 if (OpVT == MVT::f80) 212 return FPROUND_F80_F16; 213 if (OpVT == MVT::f128) 214 return FPROUND_F128_F16; 215 if (OpVT == MVT::ppcf128) 216 return FPROUND_PPCF128_F16; 217 } else if (RetVT == MVT::f32) { 218 if (OpVT == MVT::f64) 219 return FPROUND_F64_F32; 220 if (OpVT == MVT::f80) 221 return FPROUND_F80_F32; 222 if (OpVT == MVT::f128) 223 return FPROUND_F128_F32; 224 if (OpVT == MVT::ppcf128) 225 return FPROUND_PPCF128_F32; 226 } else if (RetVT == MVT::f64) { 227 if (OpVT == MVT::f80) 228 return FPROUND_F80_F64; 229 if (OpVT == MVT::f128) 230 return FPROUND_F128_F64; 231 if (OpVT == MVT::ppcf128) 232 return FPROUND_PPCF128_F64; 233 } else if (RetVT == MVT::f80) { 234 if (OpVT == MVT::f128) 235 return FPROUND_F128_F80; 236 } 237 238 return UNKNOWN_LIBCALL; 239 } 240 241 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 242 /// UNKNOWN_LIBCALL if there is none. 243 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 244 if (OpVT == MVT::f32) { 245 if (RetVT == MVT::i32) 246 return FPTOSINT_F32_I32; 247 if (RetVT == MVT::i64) 248 return FPTOSINT_F32_I64; 249 if (RetVT == MVT::i128) 250 return FPTOSINT_F32_I128; 251 } else if (OpVT == MVT::f64) { 252 if (RetVT == MVT::i32) 253 return FPTOSINT_F64_I32; 254 if (RetVT == MVT::i64) 255 return FPTOSINT_F64_I64; 256 if (RetVT == MVT::i128) 257 return FPTOSINT_F64_I128; 258 } else if (OpVT == MVT::f80) { 259 if (RetVT == MVT::i32) 260 return FPTOSINT_F80_I32; 261 if (RetVT == MVT::i64) 262 return FPTOSINT_F80_I64; 263 if (RetVT == MVT::i128) 264 return FPTOSINT_F80_I128; 265 } else if (OpVT == MVT::f128) { 266 if (RetVT == MVT::i32) 267 return FPTOSINT_F128_I32; 268 if (RetVT == MVT::i64) 269 return FPTOSINT_F128_I64; 270 if (RetVT == MVT::i128) 271 return FPTOSINT_F128_I128; 272 } else if (OpVT == MVT::ppcf128) { 273 if (RetVT == MVT::i32) 274 return FPTOSINT_PPCF128_I32; 275 if (RetVT == MVT::i64) 276 return FPTOSINT_PPCF128_I64; 277 if (RetVT == MVT::i128) 278 return FPTOSINT_PPCF128_I128; 279 } 280 return UNKNOWN_LIBCALL; 281 } 282 283 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 284 /// UNKNOWN_LIBCALL if there is none. 285 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 286 if (OpVT == MVT::f32) { 287 if (RetVT == MVT::i32) 288 return FPTOUINT_F32_I32; 289 if (RetVT == MVT::i64) 290 return FPTOUINT_F32_I64; 291 if (RetVT == MVT::i128) 292 return FPTOUINT_F32_I128; 293 } else if (OpVT == MVT::f64) { 294 if (RetVT == MVT::i32) 295 return FPTOUINT_F64_I32; 296 if (RetVT == MVT::i64) 297 return FPTOUINT_F64_I64; 298 if (RetVT == MVT::i128) 299 return FPTOUINT_F64_I128; 300 } else if (OpVT == MVT::f80) { 301 if (RetVT == MVT::i32) 302 return FPTOUINT_F80_I32; 303 if (RetVT == MVT::i64) 304 return FPTOUINT_F80_I64; 305 if (RetVT == MVT::i128) 306 return FPTOUINT_F80_I128; 307 } else if (OpVT == MVT::f128) { 308 if (RetVT == MVT::i32) 309 return FPTOUINT_F128_I32; 310 if (RetVT == MVT::i64) 311 return FPTOUINT_F128_I64; 312 if (RetVT == MVT::i128) 313 return FPTOUINT_F128_I128; 314 } else if (OpVT == MVT::ppcf128) { 315 if (RetVT == MVT::i32) 316 return FPTOUINT_PPCF128_I32; 317 if (RetVT == MVT::i64) 318 return FPTOUINT_PPCF128_I64; 319 if (RetVT == MVT::i128) 320 return FPTOUINT_PPCF128_I128; 321 } 322 return UNKNOWN_LIBCALL; 323 } 324 325 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 326 /// UNKNOWN_LIBCALL if there is none. 327 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 328 if (OpVT == MVT::i32) { 329 if (RetVT == MVT::f32) 330 return SINTTOFP_I32_F32; 331 if (RetVT == MVT::f64) 332 return SINTTOFP_I32_F64; 333 if (RetVT == MVT::f80) 334 return SINTTOFP_I32_F80; 335 if (RetVT == MVT::f128) 336 return SINTTOFP_I32_F128; 337 if (RetVT == MVT::ppcf128) 338 return SINTTOFP_I32_PPCF128; 339 } else if (OpVT == MVT::i64) { 340 if (RetVT == MVT::f32) 341 return SINTTOFP_I64_F32; 342 if (RetVT == MVT::f64) 343 return SINTTOFP_I64_F64; 344 if (RetVT == MVT::f80) 345 return SINTTOFP_I64_F80; 346 if (RetVT == MVT::f128) 347 return SINTTOFP_I64_F128; 348 if (RetVT == MVT::ppcf128) 349 return SINTTOFP_I64_PPCF128; 350 } else if (OpVT == MVT::i128) { 351 if (RetVT == MVT::f32) 352 return SINTTOFP_I128_F32; 353 if (RetVT == MVT::f64) 354 return SINTTOFP_I128_F64; 355 if (RetVT == MVT::f80) 356 return SINTTOFP_I128_F80; 357 if (RetVT == MVT::f128) 358 return SINTTOFP_I128_F128; 359 if (RetVT == MVT::ppcf128) 360 return SINTTOFP_I128_PPCF128; 361 } 362 return UNKNOWN_LIBCALL; 363 } 364 365 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 366 /// UNKNOWN_LIBCALL if there is none. 367 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 368 if (OpVT == MVT::i32) { 369 if (RetVT == MVT::f32) 370 return UINTTOFP_I32_F32; 371 if (RetVT == MVT::f64) 372 return UINTTOFP_I32_F64; 373 if (RetVT == MVT::f80) 374 return UINTTOFP_I32_F80; 375 if (RetVT == MVT::f128) 376 return UINTTOFP_I32_F128; 377 if (RetVT == MVT::ppcf128) 378 return UINTTOFP_I32_PPCF128; 379 } else if (OpVT == MVT::i64) { 380 if (RetVT == MVT::f32) 381 return UINTTOFP_I64_F32; 382 if (RetVT == MVT::f64) 383 return UINTTOFP_I64_F64; 384 if (RetVT == MVT::f80) 385 return UINTTOFP_I64_F80; 386 if (RetVT == MVT::f128) 387 return UINTTOFP_I64_F128; 388 if (RetVT == MVT::ppcf128) 389 return UINTTOFP_I64_PPCF128; 390 } else if (OpVT == MVT::i128) { 391 if (RetVT == MVT::f32) 392 return UINTTOFP_I128_F32; 393 if (RetVT == MVT::f64) 394 return UINTTOFP_I128_F64; 395 if (RetVT == MVT::f80) 396 return UINTTOFP_I128_F80; 397 if (RetVT == MVT::f128) 398 return UINTTOFP_I128_F128; 399 if (RetVT == MVT::ppcf128) 400 return UINTTOFP_I128_PPCF128; 401 } 402 return UNKNOWN_LIBCALL; 403 } 404 405 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 406 #define OP_TO_LIBCALL(Name, Enum) \ 407 case Name: \ 408 switch (VT.SimpleTy) { \ 409 default: \ 410 return UNKNOWN_LIBCALL; \ 411 case MVT::i8: \ 412 return Enum##_1; \ 413 case MVT::i16: \ 414 return Enum##_2; \ 415 case MVT::i32: \ 416 return Enum##_4; \ 417 case MVT::i64: \ 418 return Enum##_8; \ 419 case MVT::i128: \ 420 return Enum##_16; \ 421 } 422 423 switch (Opc) { 424 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 425 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 426 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 427 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 428 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 429 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 430 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 431 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 432 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 433 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 434 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 435 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 436 } 437 438 #undef OP_TO_LIBCALL 439 440 return UNKNOWN_LIBCALL; 441 } 442 443 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 444 switch (ElementSize) { 445 case 1: 446 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 447 case 2: 448 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 449 case 4: 450 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 451 case 8: 452 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 453 case 16: 454 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 455 default: 456 return UNKNOWN_LIBCALL; 457 } 458 } 459 460 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 461 switch (ElementSize) { 462 case 1: 463 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 464 case 2: 465 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 466 case 4: 467 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 468 case 8: 469 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 470 case 16: 471 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 472 default: 473 return UNKNOWN_LIBCALL; 474 } 475 } 476 477 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 478 switch (ElementSize) { 479 case 1: 480 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 481 case 2: 482 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 483 case 4: 484 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 485 case 8: 486 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 487 case 16: 488 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 489 default: 490 return UNKNOWN_LIBCALL; 491 } 492 } 493 494 /// InitCmpLibcallCCs - Set default comparison libcall CC. 495 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 496 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 497 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 498 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 499 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 500 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 501 CCs[RTLIB::UNE_F32] = ISD::SETNE; 502 CCs[RTLIB::UNE_F64] = ISD::SETNE; 503 CCs[RTLIB::UNE_F128] = ISD::SETNE; 504 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 505 CCs[RTLIB::OGE_F32] = ISD::SETGE; 506 CCs[RTLIB::OGE_F64] = ISD::SETGE; 507 CCs[RTLIB::OGE_F128] = ISD::SETGE; 508 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 509 CCs[RTLIB::OLT_F32] = ISD::SETLT; 510 CCs[RTLIB::OLT_F64] = ISD::SETLT; 511 CCs[RTLIB::OLT_F128] = ISD::SETLT; 512 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 513 CCs[RTLIB::OLE_F32] = ISD::SETLE; 514 CCs[RTLIB::OLE_F64] = ISD::SETLE; 515 CCs[RTLIB::OLE_F128] = ISD::SETLE; 516 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 517 CCs[RTLIB::OGT_F32] = ISD::SETGT; 518 CCs[RTLIB::OGT_F64] = ISD::SETGT; 519 CCs[RTLIB::OGT_F128] = ISD::SETGT; 520 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 521 CCs[RTLIB::UO_F32] = ISD::SETNE; 522 CCs[RTLIB::UO_F64] = ISD::SETNE; 523 CCs[RTLIB::UO_F128] = ISD::SETNE; 524 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 525 CCs[RTLIB::O_F32] = ISD::SETEQ; 526 CCs[RTLIB::O_F64] = ISD::SETEQ; 527 CCs[RTLIB::O_F128] = ISD::SETEQ; 528 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 529 } 530 531 /// NOTE: The TargetMachine owns TLOF. 532 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 533 initActions(); 534 535 // Perform these initializations only once. 536 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 537 MaxLoadsPerMemcmp = 8; 538 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 539 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 540 UseUnderscoreSetJmp = false; 541 UseUnderscoreLongJmp = false; 542 HasMultipleConditionRegisters = false; 543 HasExtractBitsInsn = false; 544 JumpIsExpensive = JumpIsExpensiveOverride; 545 PredictableSelectIsExpensive = false; 546 EnableExtLdPromotion = false; 547 HasFloatingPointExceptions = true; 548 StackPointerRegisterToSaveRestore = 0; 549 BooleanContents = UndefinedBooleanContent; 550 BooleanFloatContents = UndefinedBooleanContent; 551 BooleanVectorContents = UndefinedBooleanContent; 552 SchedPreferenceInfo = Sched::ILP; 553 JumpBufSize = 0; 554 JumpBufAlignment = 0; 555 MinFunctionAlignment = 0; 556 PrefFunctionAlignment = 0; 557 PrefLoopAlignment = 0; 558 GatherAllAliasesMaxDepth = 18; 559 MinStackArgumentAlignment = 1; 560 // TODO: the default will be switched to 0 in the next commit, along 561 // with the Target-specific changes necessary. 562 MaxAtomicSizeInBitsSupported = 1024; 563 564 MinCmpXchgSizeInBits = 0; 565 SupportsUnalignedAtomics = false; 566 567 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 568 569 InitLibcalls(TM.getTargetTriple()); 570 InitCmpLibcallCCs(CmpLibcallCCs); 571 } 572 573 void TargetLoweringBase::initActions() { 574 // All operations default to being supported. 575 memset(OpActions, 0, sizeof(OpActions)); 576 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 577 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 578 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 579 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 580 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 581 std::fill(std::begin(TargetDAGCombineArray), 582 std::end(TargetDAGCombineArray), 0); 583 584 // Set default actions for various operations. 585 for (MVT VT : MVT::all_valuetypes()) { 586 // Default all indexed load / store to expand. 587 for (unsigned IM = (unsigned)ISD::PRE_INC; 588 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 589 setIndexedLoadAction(IM, VT, Expand); 590 setIndexedStoreAction(IM, VT, Expand); 591 } 592 593 // Most backends expect to see the node which just returns the value loaded. 594 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 595 596 // These operations default to expand. 597 setOperationAction(ISD::FGETSIGN, VT, Expand); 598 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 599 setOperationAction(ISD::FMINNUM, VT, Expand); 600 setOperationAction(ISD::FMAXNUM, VT, Expand); 601 setOperationAction(ISD::FMINNAN, VT, Expand); 602 setOperationAction(ISD::FMAXNAN, VT, Expand); 603 setOperationAction(ISD::FMAD, VT, Expand); 604 setOperationAction(ISD::SMIN, VT, Expand); 605 setOperationAction(ISD::SMAX, VT, Expand); 606 setOperationAction(ISD::UMIN, VT, Expand); 607 setOperationAction(ISD::UMAX, VT, Expand); 608 setOperationAction(ISD::ABS, VT, Expand); 609 610 // Overflow operations default to expand 611 setOperationAction(ISD::SADDO, VT, Expand); 612 setOperationAction(ISD::SSUBO, VT, Expand); 613 setOperationAction(ISD::UADDO, VT, Expand); 614 setOperationAction(ISD::USUBO, VT, Expand); 615 setOperationAction(ISD::SMULO, VT, Expand); 616 setOperationAction(ISD::UMULO, VT, Expand); 617 618 // ADDCARRY operations default to expand 619 setOperationAction(ISD::ADDCARRY, VT, Expand); 620 setOperationAction(ISD::SUBCARRY, VT, Expand); 621 setOperationAction(ISD::SETCCCARRY, VT, Expand); 622 623 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 624 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 625 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 626 627 setOperationAction(ISD::BITREVERSE, VT, Expand); 628 629 // These library functions default to expand. 630 setOperationAction(ISD::FROUND, VT, Expand); 631 setOperationAction(ISD::FPOWI, VT, Expand); 632 633 // These operations default to expand for vector types. 634 if (VT.isVector()) { 635 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 636 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 637 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 638 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 639 } 640 641 // For most targets @llvm.get.dynamic.area.offset just returns 0. 642 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 643 } 644 645 // Most targets ignore the @llvm.prefetch intrinsic. 646 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 647 648 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 649 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 650 651 // ConstantFP nodes default to expand. Targets can either change this to 652 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 653 // to optimize expansions for certain constants. 654 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 655 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 656 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 657 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 658 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 659 660 // These library functions default to expand. 661 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 662 setOperationAction(ISD::FLOG , VT, Expand); 663 setOperationAction(ISD::FLOG2, VT, Expand); 664 setOperationAction(ISD::FLOG10, VT, Expand); 665 setOperationAction(ISD::FEXP , VT, Expand); 666 setOperationAction(ISD::FEXP2, VT, Expand); 667 setOperationAction(ISD::FFLOOR, VT, Expand); 668 setOperationAction(ISD::FNEARBYINT, VT, Expand); 669 setOperationAction(ISD::FCEIL, VT, Expand); 670 setOperationAction(ISD::FRINT, VT, Expand); 671 setOperationAction(ISD::FTRUNC, VT, Expand); 672 setOperationAction(ISD::FROUND, VT, Expand); 673 } 674 675 // Default ISD::TRAP to expand (which turns it into abort). 676 setOperationAction(ISD::TRAP, MVT::Other, Expand); 677 678 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 679 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 680 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 681 } 682 683 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 684 EVT) const { 685 return MVT::getIntegerVT(8 * DL.getPointerSize(0)); 686 } 687 688 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, 689 const DataLayout &DL) const { 690 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 691 if (LHSTy.isVector()) 692 return LHSTy; 693 return getScalarShiftAmountTy(DL, LHSTy); 694 } 695 696 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 697 assert(isTypeLegal(VT)); 698 switch (Op) { 699 default: 700 return false; 701 case ISD::SDIV: 702 case ISD::UDIV: 703 case ISD::SREM: 704 case ISD::UREM: 705 return true; 706 } 707 } 708 709 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 710 // If the command-line option was specified, ignore this request. 711 if (!JumpIsExpensiveOverride.getNumOccurrences()) 712 JumpIsExpensive = isExpensive; 713 } 714 715 TargetLoweringBase::LegalizeKind 716 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 717 // If this is a simple type, use the ComputeRegisterProp mechanism. 718 if (VT.isSimple()) { 719 MVT SVT = VT.getSimpleVT(); 720 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 721 MVT NVT = TransformToType[SVT.SimpleTy]; 722 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 723 724 assert((LA == TypeLegal || LA == TypeSoftenFloat || 725 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) && 726 "Promote may not follow Expand or Promote"); 727 728 if (LA == TypeSplitVector) 729 return LegalizeKind(LA, 730 EVT::getVectorVT(Context, SVT.getVectorElementType(), 731 SVT.getVectorNumElements() / 2)); 732 if (LA == TypeScalarizeVector) 733 return LegalizeKind(LA, SVT.getVectorElementType()); 734 return LegalizeKind(LA, NVT); 735 } 736 737 // Handle Extended Scalar Types. 738 if (!VT.isVector()) { 739 assert(VT.isInteger() && "Float types must be simple"); 740 unsigned BitSize = VT.getSizeInBits(); 741 // First promote to a power-of-two size, then expand if necessary. 742 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 743 EVT NVT = VT.getRoundIntegerType(Context); 744 assert(NVT != VT && "Unable to round integer VT"); 745 LegalizeKind NextStep = getTypeConversion(Context, NVT); 746 // Avoid multi-step promotion. 747 if (NextStep.first == TypePromoteInteger) 748 return NextStep; 749 // Return rounded integer type. 750 return LegalizeKind(TypePromoteInteger, NVT); 751 } 752 753 return LegalizeKind(TypeExpandInteger, 754 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 755 } 756 757 // Handle vector types. 758 unsigned NumElts = VT.getVectorNumElements(); 759 EVT EltVT = VT.getVectorElementType(); 760 761 // Vectors with only one element are always scalarized. 762 if (NumElts == 1) 763 return LegalizeKind(TypeScalarizeVector, EltVT); 764 765 // Try to widen vector elements until the element type is a power of two and 766 // promote it to a legal type later on, for example: 767 // <3 x i8> -> <4 x i8> -> <4 x i32> 768 if (EltVT.isInteger()) { 769 // Vectors with a number of elements that is not a power of two are always 770 // widened, for example <3 x i8> -> <4 x i8>. 771 if (!VT.isPow2VectorType()) { 772 NumElts = (unsigned)NextPowerOf2(NumElts); 773 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 774 return LegalizeKind(TypeWidenVector, NVT); 775 } 776 777 // Examine the element type. 778 LegalizeKind LK = getTypeConversion(Context, EltVT); 779 780 // If type is to be expanded, split the vector. 781 // <4 x i140> -> <2 x i140> 782 if (LK.first == TypeExpandInteger) 783 return LegalizeKind(TypeSplitVector, 784 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 785 786 // Promote the integer element types until a legal vector type is found 787 // or until the element integer type is too big. If a legal type was not 788 // found, fallback to the usual mechanism of widening/splitting the 789 // vector. 790 EVT OldEltVT = EltVT; 791 while (true) { 792 // Increase the bitwidth of the element to the next pow-of-two 793 // (which is greater than 8 bits). 794 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 795 .getRoundIntegerType(Context); 796 797 // Stop trying when getting a non-simple element type. 798 // Note that vector elements may be greater than legal vector element 799 // types. Example: X86 XMM registers hold 64bit element on 32bit 800 // systems. 801 if (!EltVT.isSimple()) 802 break; 803 804 // Build a new vector type and check if it is legal. 805 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 806 // Found a legal promoted vector type. 807 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 808 return LegalizeKind(TypePromoteInteger, 809 EVT::getVectorVT(Context, EltVT, NumElts)); 810 } 811 812 // Reset the type to the unexpanded type if we did not find a legal vector 813 // type with a promoted vector element type. 814 EltVT = OldEltVT; 815 } 816 817 // Try to widen the vector until a legal type is found. 818 // If there is no wider legal type, split the vector. 819 while (true) { 820 // Round up to the next power of 2. 821 NumElts = (unsigned)NextPowerOf2(NumElts); 822 823 // If there is no simple vector type with this many elements then there 824 // cannot be a larger legal vector type. Note that this assumes that 825 // there are no skipped intermediate vector types in the simple types. 826 if (!EltVT.isSimple()) 827 break; 828 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 829 if (LargerVector == MVT()) 830 break; 831 832 // If this type is legal then widen the vector. 833 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 834 return LegalizeKind(TypeWidenVector, LargerVector); 835 } 836 837 // Widen odd vectors to next power of two. 838 if (!VT.isPow2VectorType()) { 839 EVT NVT = VT.getPow2VectorType(Context); 840 return LegalizeKind(TypeWidenVector, NVT); 841 } 842 843 // Vectors with illegal element types are expanded. 844 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 845 return LegalizeKind(TypeSplitVector, NVT); 846 } 847 848 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 849 unsigned &NumIntermediates, 850 MVT &RegisterVT, 851 TargetLoweringBase *TLI) { 852 // Figure out the right, legal destination reg to copy into. 853 unsigned NumElts = VT.getVectorNumElements(); 854 MVT EltTy = VT.getVectorElementType(); 855 856 unsigned NumVectorRegs = 1; 857 858 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 859 // could break down into LHS/RHS like LegalizeDAG does. 860 if (!isPowerOf2_32(NumElts)) { 861 NumVectorRegs = NumElts; 862 NumElts = 1; 863 } 864 865 // Divide the input until we get to a supported size. This will always 866 // end with a scalar if the target doesn't support vectors. 867 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 868 NumElts >>= 1; 869 NumVectorRegs <<= 1; 870 } 871 872 NumIntermediates = NumVectorRegs; 873 874 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 875 if (!TLI->isTypeLegal(NewVT)) 876 NewVT = EltTy; 877 IntermediateVT = NewVT; 878 879 unsigned NewVTSize = NewVT.getSizeInBits(); 880 881 // Convert sizes such as i33 to i64. 882 if (!isPowerOf2_32(NewVTSize)) 883 NewVTSize = NextPowerOf2(NewVTSize); 884 885 MVT DestVT = TLI->getRegisterType(NewVT); 886 RegisterVT = DestVT; 887 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 888 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 889 890 // Otherwise, promotion or legal types use the same number of registers as 891 // the vector decimated to the appropriate level. 892 return NumVectorRegs; 893 } 894 895 /// isLegalRC - Return true if the value types that can be represented by the 896 /// specified register class are all legal. 897 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 898 const TargetRegisterClass &RC) const { 899 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 900 if (isTypeLegal(*I)) 901 return true; 902 return false; 903 } 904 905 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 906 /// sequence of memory operands that is recognized by PrologEpilogInserter. 907 MachineBasicBlock * 908 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 909 MachineBasicBlock *MBB) const { 910 MachineInstr *MI = &InitialMI; 911 MachineFunction &MF = *MI->getMF(); 912 MachineFrameInfo &MFI = MF.getFrameInfo(); 913 914 // We're handling multiple types of operands here: 915 // PATCHPOINT MetaArgs - live-in, read only, direct 916 // STATEPOINT Deopt Spill - live-through, read only, indirect 917 // STATEPOINT Deopt Alloca - live-through, read only, direct 918 // (We're currently conservative and mark the deopt slots read/write in 919 // practice.) 920 // STATEPOINT GC Spill - live-through, read/write, indirect 921 // STATEPOINT GC Alloca - live-through, read/write, direct 922 // The live-in vs live-through is handled already (the live through ones are 923 // all stack slots), but we need to handle the different type of stackmap 924 // operands and memory effects here. 925 926 // MI changes inside this loop as we grow operands. 927 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 928 MachineOperand &MO = MI->getOperand(OperIdx); 929 if (!MO.isFI()) 930 continue; 931 932 // foldMemoryOperand builds a new MI after replacing a single FI operand 933 // with the canonical set of five x86 addressing-mode operands. 934 int FI = MO.getIndex(); 935 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 936 937 // Copy operands before the frame-index. 938 for (unsigned i = 0; i < OperIdx; ++i) 939 MIB.add(MI->getOperand(i)); 940 // Add frame index operands recognized by stackmaps.cpp 941 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 942 // indirect-mem-ref tag, size, #FI, offset. 943 // Used for spills inserted by StatepointLowering. This codepath is not 944 // used for patchpoints/stackmaps at all, for these spilling is done via 945 // foldMemoryOperand callback only. 946 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 947 MIB.addImm(StackMaps::IndirectMemRefOp); 948 MIB.addImm(MFI.getObjectSize(FI)); 949 MIB.add(MI->getOperand(OperIdx)); 950 MIB.addImm(0); 951 } else { 952 // direct-mem-ref tag, #FI, offset. 953 // Used by patchpoint, and direct alloca arguments to statepoints 954 MIB.addImm(StackMaps::DirectMemRefOp); 955 MIB.add(MI->getOperand(OperIdx)); 956 MIB.addImm(0); 957 } 958 // Copy the operands after the frame index. 959 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 960 MIB.add(MI->getOperand(i)); 961 962 // Inherit previous memory operands. 963 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 964 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 965 966 // Add a new memory operand for this FI. 967 assert(MFI.getObjectOffset(FI) != -1); 968 969 auto Flags = MachineMemOperand::MOLoad; 970 if (MI->getOpcode() == TargetOpcode::STATEPOINT) { 971 Flags |= MachineMemOperand::MOStore; 972 Flags |= MachineMemOperand::MOVolatile; 973 } 974 MachineMemOperand *MMO = MF.getMachineMemOperand( 975 MachinePointerInfo::getFixedStack(MF, FI), Flags, 976 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 977 MIB->addMemOperand(MF, MMO); 978 979 // Replace the instruction and update the operand index. 980 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 981 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 982 MI->eraseFromParent(); 983 MI = MIB; 984 } 985 return MBB; 986 } 987 988 /// findRepresentativeClass - Return the largest legal super-reg register class 989 /// of the register class for the specified type and its associated "cost". 990 // This function is in TargetLowering because it uses RegClassForVT which would 991 // need to be moved to TargetRegisterInfo and would necessitate moving 992 // isTypeLegal over as well - a massive change that would just require 993 // TargetLowering having a TargetRegisterInfo class member that it would use. 994 std::pair<const TargetRegisterClass *, uint8_t> 995 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 996 MVT VT) const { 997 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 998 if (!RC) 999 return std::make_pair(RC, 0); 1000 1001 // Compute the set of all super-register classes. 1002 BitVector SuperRegRC(TRI->getNumRegClasses()); 1003 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1004 SuperRegRC.setBitsInMask(RCI.getMask()); 1005 1006 // Find the first legal register class with the largest spill size. 1007 const TargetRegisterClass *BestRC = RC; 1008 for (unsigned i : SuperRegRC.set_bits()) { 1009 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1010 // We want the largest possible spill size. 1011 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1012 continue; 1013 if (!isLegalRC(*TRI, *SuperRC)) 1014 continue; 1015 BestRC = SuperRC; 1016 } 1017 return std::make_pair(BestRC, 1); 1018 } 1019 1020 /// computeRegisterProperties - Once all of the register classes are added, 1021 /// this allows us to compute derived properties we expose. 1022 void TargetLoweringBase::computeRegisterProperties( 1023 const TargetRegisterInfo *TRI) { 1024 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1025 "Too many value types for ValueTypeActions to hold!"); 1026 1027 // Everything defaults to needing one register. 1028 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1029 NumRegistersForVT[i] = 1; 1030 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1031 } 1032 // ...except isVoid, which doesn't need any registers. 1033 NumRegistersForVT[MVT::isVoid] = 0; 1034 1035 // Find the largest integer register class. 1036 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1037 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1038 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1039 1040 // Every integer value type larger than this largest register takes twice as 1041 // many registers to represent as the previous ValueType. 1042 for (unsigned ExpandedReg = LargestIntReg + 1; 1043 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1044 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1045 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1046 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1047 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1048 TypeExpandInteger); 1049 } 1050 1051 // Inspect all of the ValueType's smaller than the largest integer 1052 // register to see which ones need promotion. 1053 unsigned LegalIntReg = LargestIntReg; 1054 for (unsigned IntReg = LargestIntReg - 1; 1055 IntReg >= (unsigned)MVT::i1; --IntReg) { 1056 MVT IVT = (MVT::SimpleValueType)IntReg; 1057 if (isTypeLegal(IVT)) { 1058 LegalIntReg = IntReg; 1059 } else { 1060 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1061 (const MVT::SimpleValueType)LegalIntReg; 1062 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1063 } 1064 } 1065 1066 // ppcf128 type is really two f64's. 1067 if (!isTypeLegal(MVT::ppcf128)) { 1068 if (isTypeLegal(MVT::f64)) { 1069 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1070 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1071 TransformToType[MVT::ppcf128] = MVT::f64; 1072 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1073 } else { 1074 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1075 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1076 TransformToType[MVT::ppcf128] = MVT::i128; 1077 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1078 } 1079 } 1080 1081 // Decide how to handle f128. If the target does not have native f128 support, 1082 // expand it to i128 and we will be generating soft float library calls. 1083 if (!isTypeLegal(MVT::f128)) { 1084 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1085 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1086 TransformToType[MVT::f128] = MVT::i128; 1087 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1088 } 1089 1090 // Decide how to handle f64. If the target does not have native f64 support, 1091 // expand it to i64 and we will be generating soft float library calls. 1092 if (!isTypeLegal(MVT::f64)) { 1093 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1094 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1095 TransformToType[MVT::f64] = MVT::i64; 1096 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1097 } 1098 1099 // Decide how to handle f32. If the target does not have native f32 support, 1100 // expand it to i32 and we will be generating soft float library calls. 1101 if (!isTypeLegal(MVT::f32)) { 1102 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1103 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1104 TransformToType[MVT::f32] = MVT::i32; 1105 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1106 } 1107 1108 // Decide how to handle f16. If the target does not have native f16 support, 1109 // promote it to f32, because there are no f16 library calls (except for 1110 // conversions). 1111 if (!isTypeLegal(MVT::f16)) { 1112 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1113 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1114 TransformToType[MVT::f16] = MVT::f32; 1115 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1116 } 1117 1118 // Loop over all of the vector value types to see which need transformations. 1119 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1120 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1121 MVT VT = (MVT::SimpleValueType) i; 1122 if (isTypeLegal(VT)) 1123 continue; 1124 1125 MVT EltVT = VT.getVectorElementType(); 1126 unsigned NElts = VT.getVectorNumElements(); 1127 bool IsLegalWiderType = false; 1128 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1129 switch (PreferredAction) { 1130 case TypePromoteInteger: 1131 // Try to promote the elements of integer vectors. If no legal 1132 // promotion was found, fall through to the widen-vector method. 1133 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1134 MVT SVT = (MVT::SimpleValueType) nVT; 1135 // Promote vectors of integers to vectors with the same number 1136 // of elements, with a wider element type. 1137 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1138 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1139 TransformToType[i] = SVT; 1140 RegisterTypeForVT[i] = SVT; 1141 NumRegistersForVT[i] = 1; 1142 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1143 IsLegalWiderType = true; 1144 break; 1145 } 1146 } 1147 if (IsLegalWiderType) 1148 break; 1149 LLVM_FALLTHROUGH; 1150 1151 case TypeWidenVector: 1152 // Try to widen the vector. 1153 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1154 MVT SVT = (MVT::SimpleValueType) nVT; 1155 if (SVT.getVectorElementType() == EltVT 1156 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1157 TransformToType[i] = SVT; 1158 RegisterTypeForVT[i] = SVT; 1159 NumRegistersForVT[i] = 1; 1160 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1161 IsLegalWiderType = true; 1162 break; 1163 } 1164 } 1165 if (IsLegalWiderType) 1166 break; 1167 LLVM_FALLTHROUGH; 1168 1169 case TypeSplitVector: 1170 case TypeScalarizeVector: { 1171 MVT IntermediateVT; 1172 MVT RegisterVT; 1173 unsigned NumIntermediates; 1174 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1175 NumIntermediates, RegisterVT, this); 1176 RegisterTypeForVT[i] = RegisterVT; 1177 1178 MVT NVT = VT.getPow2VectorType(); 1179 if (NVT == VT) { 1180 // Type is already a power of 2. The default action is to split. 1181 TransformToType[i] = MVT::Other; 1182 if (PreferredAction == TypeScalarizeVector) 1183 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1184 else if (PreferredAction == TypeSplitVector) 1185 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1186 else 1187 // Set type action according to the number of elements. 1188 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1189 : TypeSplitVector); 1190 } else { 1191 TransformToType[i] = NVT; 1192 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1193 } 1194 break; 1195 } 1196 default: 1197 llvm_unreachable("Unknown vector legalization action!"); 1198 } 1199 } 1200 1201 // Determine the 'representative' register class for each value type. 1202 // An representative register class is the largest (meaning one which is 1203 // not a sub-register class / subreg register class) legal register class for 1204 // a group of value types. For example, on i386, i8, i16, and i32 1205 // representative would be GR32; while on x86_64 it's GR64. 1206 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1207 const TargetRegisterClass* RRC; 1208 uint8_t Cost; 1209 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1210 RepRegClassForVT[i] = RRC; 1211 RepRegClassCostForVT[i] = Cost; 1212 } 1213 } 1214 1215 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1216 EVT VT) const { 1217 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1218 return getPointerTy(DL).SimpleTy; 1219 } 1220 1221 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1222 return MVT::i32; // return the default value 1223 } 1224 1225 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1226 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1227 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1228 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1229 /// 1230 /// This method returns the number of registers needed, and the VT for each 1231 /// register. It also returns the VT and quantity of the intermediate values 1232 /// before they are promoted/expanded. 1233 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1234 EVT &IntermediateVT, 1235 unsigned &NumIntermediates, 1236 MVT &RegisterVT) const { 1237 unsigned NumElts = VT.getVectorNumElements(); 1238 1239 // If there is a wider vector type with the same element type as this one, 1240 // or a promoted vector type that has the same number of elements which 1241 // are wider, then we should convert to that legal vector type. 1242 // This handles things like <2 x float> -> <4 x float> and 1243 // <4 x i1> -> <4 x i32>. 1244 LegalizeTypeAction TA = getTypeAction(Context, VT); 1245 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1246 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1247 if (isTypeLegal(RegisterEVT)) { 1248 IntermediateVT = RegisterEVT; 1249 RegisterVT = RegisterEVT.getSimpleVT(); 1250 NumIntermediates = 1; 1251 return 1; 1252 } 1253 } 1254 1255 // Figure out the right, legal destination reg to copy into. 1256 EVT EltTy = VT.getVectorElementType(); 1257 1258 unsigned NumVectorRegs = 1; 1259 1260 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1261 // could break down into LHS/RHS like LegalizeDAG does. 1262 if (!isPowerOf2_32(NumElts)) { 1263 NumVectorRegs = NumElts; 1264 NumElts = 1; 1265 } 1266 1267 // Divide the input until we get to a supported size. This will always 1268 // end with a scalar if the target doesn't support vectors. 1269 while (NumElts > 1 && !isTypeLegal( 1270 EVT::getVectorVT(Context, EltTy, NumElts))) { 1271 NumElts >>= 1; 1272 NumVectorRegs <<= 1; 1273 } 1274 1275 NumIntermediates = NumVectorRegs; 1276 1277 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1278 if (!isTypeLegal(NewVT)) 1279 NewVT = EltTy; 1280 IntermediateVT = NewVT; 1281 1282 MVT DestVT = getRegisterType(Context, NewVT); 1283 RegisterVT = DestVT; 1284 unsigned NewVTSize = NewVT.getSizeInBits(); 1285 1286 // Convert sizes such as i33 to i64. 1287 if (!isPowerOf2_32(NewVTSize)) 1288 NewVTSize = NextPowerOf2(NewVTSize); 1289 1290 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1291 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1292 1293 // Otherwise, promotion or legal types use the same number of registers as 1294 // the vector decimated to the appropriate level. 1295 return NumVectorRegs; 1296 } 1297 1298 /// Get the EVTs and ArgFlags collections that represent the legalized return 1299 /// type of the given function. This does not require a DAG or a return value, 1300 /// and is suitable for use before any DAGs for the function are constructed. 1301 /// TODO: Move this out of TargetLowering.cpp. 1302 void llvm::GetReturnInfo(Type *ReturnType, AttributeList attr, 1303 SmallVectorImpl<ISD::OutputArg> &Outs, 1304 const TargetLowering &TLI, const DataLayout &DL) { 1305 SmallVector<EVT, 4> ValueVTs; 1306 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1307 unsigned NumValues = ValueVTs.size(); 1308 if (NumValues == 0) return; 1309 1310 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1311 EVT VT = ValueVTs[j]; 1312 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1313 1314 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1315 ExtendKind = ISD::SIGN_EXTEND; 1316 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1317 ExtendKind = ISD::ZERO_EXTEND; 1318 1319 // FIXME: C calling convention requires the return type to be promoted to 1320 // at least 32-bit. But this is not necessary for non-C calling 1321 // conventions. The frontend should mark functions whose return values 1322 // require promoting with signext or zeroext attributes. 1323 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1324 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1325 if (VT.bitsLT(MinVT)) 1326 VT = MinVT; 1327 } 1328 1329 unsigned NumParts = 1330 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), VT); 1331 MVT PartVT = 1332 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), VT); 1333 1334 // 'inreg' on function refers to return value 1335 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1336 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1337 Flags.setInReg(); 1338 1339 // Propagate extension type if any 1340 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1341 Flags.setSExt(); 1342 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1343 Flags.setZExt(); 1344 1345 for (unsigned i = 0; i < NumParts; ++i) 1346 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0)); 1347 } 1348 } 1349 1350 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1351 /// function arguments in the caller parameter area. This is the actual 1352 /// alignment, not its logarithm. 1353 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1354 const DataLayout &DL) const { 1355 return DL.getABITypeAlignment(Ty); 1356 } 1357 1358 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1359 const DataLayout &DL, EVT VT, 1360 unsigned AddrSpace, 1361 unsigned Alignment, 1362 bool *Fast) const { 1363 // Check if the specified alignment is sufficient based on the data layout. 1364 // TODO: While using the data layout works in practice, a better solution 1365 // would be to implement this check directly (make this a virtual function). 1366 // For example, the ABI alignment may change based on software platform while 1367 // this function should only be affected by hardware implementation. 1368 Type *Ty = VT.getTypeForEVT(Context); 1369 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1370 // Assume that an access that meets the ABI-specified alignment is fast. 1371 if (Fast != nullptr) 1372 *Fast = true; 1373 return true; 1374 } 1375 1376 // This is a misaligned access. 1377 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast); 1378 } 1379 1380 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1381 return BranchProbability(MinPercentageForPredictableBranch, 100); 1382 } 1383 1384 //===----------------------------------------------------------------------===// 1385 // TargetTransformInfo Helpers 1386 //===----------------------------------------------------------------------===// 1387 1388 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1389 enum InstructionOpcodes { 1390 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1391 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1392 #include "llvm/IR/Instruction.def" 1393 }; 1394 switch (static_cast<InstructionOpcodes>(Opcode)) { 1395 case Ret: return 0; 1396 case Br: return 0; 1397 case Switch: return 0; 1398 case IndirectBr: return 0; 1399 case Invoke: return 0; 1400 case Resume: return 0; 1401 case Unreachable: return 0; 1402 case CleanupRet: return 0; 1403 case CatchRet: return 0; 1404 case CatchPad: return 0; 1405 case CatchSwitch: return 0; 1406 case CleanupPad: return 0; 1407 case Add: return ISD::ADD; 1408 case FAdd: return ISD::FADD; 1409 case Sub: return ISD::SUB; 1410 case FSub: return ISD::FSUB; 1411 case Mul: return ISD::MUL; 1412 case FMul: return ISD::FMUL; 1413 case UDiv: return ISD::UDIV; 1414 case SDiv: return ISD::SDIV; 1415 case FDiv: return ISD::FDIV; 1416 case URem: return ISD::UREM; 1417 case SRem: return ISD::SREM; 1418 case FRem: return ISD::FREM; 1419 case Shl: return ISD::SHL; 1420 case LShr: return ISD::SRL; 1421 case AShr: return ISD::SRA; 1422 case And: return ISD::AND; 1423 case Or: return ISD::OR; 1424 case Xor: return ISD::XOR; 1425 case Alloca: return 0; 1426 case Load: return ISD::LOAD; 1427 case Store: return ISD::STORE; 1428 case GetElementPtr: return 0; 1429 case Fence: return 0; 1430 case AtomicCmpXchg: return 0; 1431 case AtomicRMW: return 0; 1432 case Trunc: return ISD::TRUNCATE; 1433 case ZExt: return ISD::ZERO_EXTEND; 1434 case SExt: return ISD::SIGN_EXTEND; 1435 case FPToUI: return ISD::FP_TO_UINT; 1436 case FPToSI: return ISD::FP_TO_SINT; 1437 case UIToFP: return ISD::UINT_TO_FP; 1438 case SIToFP: return ISD::SINT_TO_FP; 1439 case FPTrunc: return ISD::FP_ROUND; 1440 case FPExt: return ISD::FP_EXTEND; 1441 case PtrToInt: return ISD::BITCAST; 1442 case IntToPtr: return ISD::BITCAST; 1443 case BitCast: return ISD::BITCAST; 1444 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1445 case ICmp: return ISD::SETCC; 1446 case FCmp: return ISD::SETCC; 1447 case PHI: return 0; 1448 case Call: return 0; 1449 case Select: return ISD::SELECT; 1450 case UserOp1: return 0; 1451 case UserOp2: return 0; 1452 case VAArg: return 0; 1453 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1454 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1455 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1456 case ExtractValue: return ISD::MERGE_VALUES; 1457 case InsertValue: return ISD::MERGE_VALUES; 1458 case LandingPad: return 0; 1459 } 1460 1461 llvm_unreachable("Unknown instruction type encountered!"); 1462 } 1463 1464 std::pair<int, MVT> 1465 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1466 Type *Ty) const { 1467 LLVMContext &C = Ty->getContext(); 1468 EVT MTy = getValueType(DL, Ty); 1469 1470 int Cost = 1; 1471 // We keep legalizing the type until we find a legal kind. We assume that 1472 // the only operation that costs anything is the split. After splitting 1473 // we need to handle two types. 1474 while (true) { 1475 LegalizeKind LK = getTypeConversion(C, MTy); 1476 1477 if (LK.first == TypeLegal) 1478 return std::make_pair(Cost, MTy.getSimpleVT()); 1479 1480 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1481 Cost *= 2; 1482 1483 // Do not loop with f128 type. 1484 if (MTy == LK.second) 1485 return std::make_pair(Cost, MTy.getSimpleVT()); 1486 1487 // Keep legalizing the type. 1488 MTy = LK.second; 1489 } 1490 } 1491 1492 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1493 bool UseTLS) const { 1494 // compiler-rt provides a variable with a magic name. Targets that do not 1495 // link with compiler-rt may also provide such a variable. 1496 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1497 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1498 auto UnsafeStackPtr = 1499 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1500 1501 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1502 1503 if (!UnsafeStackPtr) { 1504 auto TLSModel = UseTLS ? 1505 GlobalValue::InitialExecTLSModel : 1506 GlobalValue::NotThreadLocal; 1507 // The global variable is not defined yet, define it ourselves. 1508 // We use the initial-exec TLS model because we do not support the 1509 // variable living anywhere other than in the main executable. 1510 UnsafeStackPtr = new GlobalVariable( 1511 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1512 UnsafeStackPtrVar, nullptr, TLSModel); 1513 } else { 1514 // The variable exists, check its type and attributes. 1515 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1516 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1517 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1518 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1519 (UseTLS ? "" : "not ") + "be thread-local"); 1520 } 1521 return UnsafeStackPtr; 1522 } 1523 1524 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1525 if (!TM.getTargetTriple().isAndroid()) 1526 return getDefaultSafeStackPointerLocation(IRB, true); 1527 1528 // Android provides a libc function to retrieve the address of the current 1529 // thread's unsafe stack pointer. 1530 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1531 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1532 Value *Fn = M->getOrInsertFunction("__safestack_pointer_address", 1533 StackPtrTy->getPointerTo(0)); 1534 return IRB.CreateCall(Fn); 1535 } 1536 1537 //===----------------------------------------------------------------------===// 1538 // Loop Strength Reduction hooks 1539 //===----------------------------------------------------------------------===// 1540 1541 /// isLegalAddressingMode - Return true if the addressing mode represented 1542 /// by AM is legal for this target, for a load/store of the specified type. 1543 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1544 const AddrMode &AM, Type *Ty, 1545 unsigned AS, Instruction *I) const { 1546 // The default implementation of this implements a conservative RISCy, r+r and 1547 // r+i addr mode. 1548 1549 // Allows a sign-extended 16-bit immediate field. 1550 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1551 return false; 1552 1553 // No global is ever allowed as a base. 1554 if (AM.BaseGV) 1555 return false; 1556 1557 // Only support r+r, 1558 switch (AM.Scale) { 1559 case 0: // "r+i" or just "i", depending on HasBaseReg. 1560 break; 1561 case 1: 1562 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1563 return false; 1564 // Otherwise we have r+r or r+i. 1565 break; 1566 case 2: 1567 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1568 return false; 1569 // Allow 2*r as r+r. 1570 break; 1571 default: // Don't allow n * r 1572 return false; 1573 } 1574 1575 return true; 1576 } 1577 1578 //===----------------------------------------------------------------------===// 1579 // Stack Protector 1580 //===----------------------------------------------------------------------===// 1581 1582 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1583 // so that SelectionDAG handle SSP. 1584 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1585 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1586 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1587 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1588 return M.getOrInsertGlobal("__guard_local", PtrTy); 1589 } 1590 return nullptr; 1591 } 1592 1593 // Currently only support "standard" __stack_chk_guard. 1594 // TODO: add LOAD_STACK_GUARD support. 1595 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1596 M.getOrInsertGlobal("__stack_chk_guard", Type::getInt8PtrTy(M.getContext())); 1597 } 1598 1599 // Currently only support "standard" __stack_chk_guard. 1600 // TODO: add LOAD_STACK_GUARD support. 1601 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1602 return M.getGlobalVariable("__stack_chk_guard", true); 1603 } 1604 1605 Value *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1606 return nullptr; 1607 } 1608 1609 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1610 return MinimumJumpTableEntries; 1611 } 1612 1613 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1614 MinimumJumpTableEntries = Val; 1615 } 1616 1617 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1618 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1619 } 1620 1621 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1622 return MaximumJumpTableSize; 1623 } 1624 1625 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1626 MaximumJumpTableSize = Val; 1627 } 1628 1629 //===----------------------------------------------------------------------===// 1630 // Reciprocal Estimates 1631 //===----------------------------------------------------------------------===// 1632 1633 /// Get the reciprocal estimate attribute string for a function that will 1634 /// override the target defaults. 1635 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1636 const Function &F = MF.getFunction(); 1637 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1638 } 1639 1640 /// Construct a string for the given reciprocal operation of the given type. 1641 /// This string should match the corresponding option to the front-end's 1642 /// "-mrecip" flag assuming those strings have been passed through in an 1643 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1644 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1645 std::string Name = VT.isVector() ? "vec-" : ""; 1646 1647 Name += IsSqrt ? "sqrt" : "div"; 1648 1649 // TODO: Handle "half" or other float types? 1650 if (VT.getScalarType() == MVT::f64) { 1651 Name += "d"; 1652 } else { 1653 assert(VT.getScalarType() == MVT::f32 && 1654 "Unexpected FP type for reciprocal estimate"); 1655 Name += "f"; 1656 } 1657 1658 return Name; 1659 } 1660 1661 /// Return the character position and value (a single numeric character) of a 1662 /// customized refinement operation in the input string if it exists. Return 1663 /// false if there is no customized refinement step count. 1664 static bool parseRefinementStep(StringRef In, size_t &Position, 1665 uint8_t &Value) { 1666 const char RefStepToken = ':'; 1667 Position = In.find(RefStepToken); 1668 if (Position == StringRef::npos) 1669 return false; 1670 1671 StringRef RefStepString = In.substr(Position + 1); 1672 // Allow exactly one numeric character for the additional refinement 1673 // step parameter. 1674 if (RefStepString.size() == 1) { 1675 char RefStepChar = RefStepString[0]; 1676 if (RefStepChar >= '0' && RefStepChar <= '9') { 1677 Value = RefStepChar - '0'; 1678 return true; 1679 } 1680 } 1681 report_fatal_error("Invalid refinement step for -recip."); 1682 } 1683 1684 /// For the input attribute string, return one of the ReciprocalEstimate enum 1685 /// status values (enabled, disabled, or not specified) for this operation on 1686 /// the specified data type. 1687 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1688 if (Override.empty()) 1689 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1690 1691 SmallVector<StringRef, 4> OverrideVector; 1692 SplitString(Override, OverrideVector, ","); 1693 unsigned NumArgs = OverrideVector.size(); 1694 1695 // Check if "all", "none", or "default" was specified. 1696 if (NumArgs == 1) { 1697 // Look for an optional setting of the number of refinement steps needed 1698 // for this type of reciprocal operation. 1699 size_t RefPos; 1700 uint8_t RefSteps; 1701 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1702 // Split the string for further processing. 1703 Override = Override.substr(0, RefPos); 1704 } 1705 1706 // All reciprocal types are enabled. 1707 if (Override == "all") 1708 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1709 1710 // All reciprocal types are disabled. 1711 if (Override == "none") 1712 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1713 1714 // Target defaults for enablement are used. 1715 if (Override == "default") 1716 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1717 } 1718 1719 // The attribute string may omit the size suffix ('f'/'d'). 1720 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1721 std::string VTNameNoSize = VTName; 1722 VTNameNoSize.pop_back(); 1723 static const char DisabledPrefix = '!'; 1724 1725 for (StringRef RecipType : OverrideVector) { 1726 size_t RefPos; 1727 uint8_t RefSteps; 1728 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1729 RecipType = RecipType.substr(0, RefPos); 1730 1731 // Ignore the disablement token for string matching. 1732 bool IsDisabled = RecipType[0] == DisabledPrefix; 1733 if (IsDisabled) 1734 RecipType = RecipType.substr(1); 1735 1736 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1737 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1738 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1739 } 1740 1741 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1742 } 1743 1744 /// For the input attribute string, return the customized refinement step count 1745 /// for this operation on the specified data type. If the step count does not 1746 /// exist, return the ReciprocalEstimate enum value for unspecified. 1747 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1748 if (Override.empty()) 1749 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1750 1751 SmallVector<StringRef, 4> OverrideVector; 1752 SplitString(Override, OverrideVector, ","); 1753 unsigned NumArgs = OverrideVector.size(); 1754 1755 // Check if "all", "default", or "none" was specified. 1756 if (NumArgs == 1) { 1757 // Look for an optional setting of the number of refinement steps needed 1758 // for this type of reciprocal operation. 1759 size_t RefPos; 1760 uint8_t RefSteps; 1761 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1762 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1763 1764 // Split the string for further processing. 1765 Override = Override.substr(0, RefPos); 1766 assert(Override != "none" && 1767 "Disabled reciprocals, but specifed refinement steps?"); 1768 1769 // If this is a general override, return the specified number of steps. 1770 if (Override == "all" || Override == "default") 1771 return RefSteps; 1772 } 1773 1774 // The attribute string may omit the size suffix ('f'/'d'). 1775 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1776 std::string VTNameNoSize = VTName; 1777 VTNameNoSize.pop_back(); 1778 1779 for (StringRef RecipType : OverrideVector) { 1780 size_t RefPos; 1781 uint8_t RefSteps; 1782 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1783 continue; 1784 1785 RecipType = RecipType.substr(0, RefPos); 1786 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1787 return RefSteps; 1788 } 1789 1790 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1791 } 1792 1793 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1794 MachineFunction &MF) const { 1795 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1796 } 1797 1798 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1799 MachineFunction &MF) const { 1800 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1801 } 1802 1803 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1804 MachineFunction &MF) const { 1805 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1806 } 1807 1808 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1809 MachineFunction &MF) const { 1810 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1811 } 1812 1813 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1814 MF.getRegInfo().freezeReservedRegs(MF); 1815 } 1816