1 //===- TargetLoweringBase.cpp - Implement the TargetLoweringBase class ----===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This implements the TargetLoweringBase class. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/ADT/BitVector.h" 14 #include "llvm/ADT/STLExtras.h" 15 #include "llvm/ADT/SmallVector.h" 16 #include "llvm/ADT/StringExtras.h" 17 #include "llvm/ADT/StringRef.h" 18 #include "llvm/ADT/Triple.h" 19 #include "llvm/ADT/Twine.h" 20 #include "llvm/CodeGen/Analysis.h" 21 #include "llvm/CodeGen/ISDOpcodes.h" 22 #include "llvm/CodeGen/MachineBasicBlock.h" 23 #include "llvm/CodeGen/MachineFrameInfo.h" 24 #include "llvm/CodeGen/MachineFunction.h" 25 #include "llvm/CodeGen/MachineInstr.h" 26 #include "llvm/CodeGen/MachineInstrBuilder.h" 27 #include "llvm/CodeGen/MachineMemOperand.h" 28 #include "llvm/CodeGen/MachineOperand.h" 29 #include "llvm/CodeGen/MachineRegisterInfo.h" 30 #include "llvm/CodeGen/RuntimeLibcalls.h" 31 #include "llvm/CodeGen/StackMaps.h" 32 #include "llvm/CodeGen/TargetLowering.h" 33 #include "llvm/CodeGen/TargetOpcodes.h" 34 #include "llvm/CodeGen/TargetRegisterInfo.h" 35 #include "llvm/CodeGen/ValueTypes.h" 36 #include "llvm/IR/Attributes.h" 37 #include "llvm/IR/CallingConv.h" 38 #include "llvm/IR/DataLayout.h" 39 #include "llvm/IR/DerivedTypes.h" 40 #include "llvm/IR/Function.h" 41 #include "llvm/IR/GlobalValue.h" 42 #include "llvm/IR/GlobalVariable.h" 43 #include "llvm/IR/IRBuilder.h" 44 #include "llvm/IR/Module.h" 45 #include "llvm/IR/Type.h" 46 #include "llvm/Support/BranchProbability.h" 47 #include "llvm/Support/Casting.h" 48 #include "llvm/Support/CommandLine.h" 49 #include "llvm/Support/Compiler.h" 50 #include "llvm/Support/ErrorHandling.h" 51 #include "llvm/Support/MachineValueType.h" 52 #include "llvm/Support/MathExtras.h" 53 #include "llvm/Target/TargetMachine.h" 54 #include <algorithm> 55 #include <cassert> 56 #include <cstddef> 57 #include <cstdint> 58 #include <cstring> 59 #include <iterator> 60 #include <string> 61 #include <tuple> 62 #include <utility> 63 64 using namespace llvm; 65 66 static cl::opt<bool> JumpIsExpensiveOverride( 67 "jump-is-expensive", cl::init(false), 68 cl::desc("Do not create extra branches to split comparison logic."), 69 cl::Hidden); 70 71 static cl::opt<unsigned> MinimumJumpTableEntries 72 ("min-jump-table-entries", cl::init(4), cl::Hidden, 73 cl::desc("Set minimum number of entries to use a jump table.")); 74 75 static cl::opt<unsigned> MaximumJumpTableSize 76 ("max-jump-table-size", cl::init(0), cl::Hidden, 77 cl::desc("Set maximum size of jump tables; zero for no limit.")); 78 79 /// Minimum jump table density for normal functions. 80 static cl::opt<unsigned> 81 JumpTableDensity("jump-table-density", cl::init(10), cl::Hidden, 82 cl::desc("Minimum density for building a jump table in " 83 "a normal function")); 84 85 /// Minimum jump table density for -Os or -Oz functions. 86 static cl::opt<unsigned> OptsizeJumpTableDensity( 87 "optsize-jump-table-density", cl::init(40), cl::Hidden, 88 cl::desc("Minimum density for building a jump table in " 89 "an optsize function")); 90 91 static bool darwinHasSinCos(const Triple &TT) { 92 assert(TT.isOSDarwin() && "should be called with darwin triple"); 93 // Don't bother with 32 bit x86. 94 if (TT.getArch() == Triple::x86) 95 return false; 96 // Macos < 10.9 has no sincos_stret. 97 if (TT.isMacOSX()) 98 return !TT.isMacOSXVersionLT(10, 9) && TT.isArch64Bit(); 99 // iOS < 7.0 has no sincos_stret. 100 if (TT.isiOS()) 101 return !TT.isOSVersionLT(7, 0); 102 // Any other darwin such as WatchOS/TvOS is new enough. 103 return true; 104 } 105 106 // Although this default value is arbitrary, it is not random. It is assumed 107 // that a condition that evaluates the same way by a higher percentage than this 108 // is best represented as control flow. Therefore, the default value N should be 109 // set such that the win from N% correct executions is greater than the loss 110 // from (100 - N)% mispredicted executions for the majority of intended targets. 111 static cl::opt<int> MinPercentageForPredictableBranch( 112 "min-predictable-branch", cl::init(99), 113 cl::desc("Minimum percentage (0-100) that a condition must be either true " 114 "or false to assume that the condition is predictable"), 115 cl::Hidden); 116 117 void TargetLoweringBase::InitLibcalls(const Triple &TT) { 118 #define HANDLE_LIBCALL(code, name) \ 119 setLibcallName(RTLIB::code, name); 120 #include "llvm/IR/RuntimeLibcalls.def" 121 #undef HANDLE_LIBCALL 122 // Initialize calling conventions to their default. 123 for (int LC = 0; LC < RTLIB::UNKNOWN_LIBCALL; ++LC) 124 setLibcallCallingConv((RTLIB::Libcall)LC, CallingConv::C); 125 126 // A few names are different on particular architectures or environments. 127 if (TT.isOSDarwin()) { 128 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead 129 // of the gnueabi-style __gnu_*_ieee. 130 // FIXME: What about other targets? 131 setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2"); 132 setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2"); 133 134 // Some darwins have an optimized __bzero/bzero function. 135 switch (TT.getArch()) { 136 case Triple::x86: 137 case Triple::x86_64: 138 if (TT.isMacOSX() && !TT.isMacOSXVersionLT(10, 6)) 139 setLibcallName(RTLIB::BZERO, "__bzero"); 140 break; 141 case Triple::aarch64: 142 setLibcallName(RTLIB::BZERO, "bzero"); 143 break; 144 default: 145 break; 146 } 147 148 if (darwinHasSinCos(TT)) { 149 setLibcallName(RTLIB::SINCOS_STRET_F32, "__sincosf_stret"); 150 setLibcallName(RTLIB::SINCOS_STRET_F64, "__sincos_stret"); 151 if (TT.isWatchABI()) { 152 setLibcallCallingConv(RTLIB::SINCOS_STRET_F32, 153 CallingConv::ARM_AAPCS_VFP); 154 setLibcallCallingConv(RTLIB::SINCOS_STRET_F64, 155 CallingConv::ARM_AAPCS_VFP); 156 } 157 } 158 } else { 159 setLibcallName(RTLIB::FPEXT_F16_F32, "__gnu_h2f_ieee"); 160 setLibcallName(RTLIB::FPROUND_F32_F16, "__gnu_f2h_ieee"); 161 } 162 163 if (TT.isGNUEnvironment() || TT.isOSFuchsia() || 164 (TT.isAndroid() && !TT.isAndroidVersionLT(9))) { 165 setLibcallName(RTLIB::SINCOS_F32, "sincosf"); 166 setLibcallName(RTLIB::SINCOS_F64, "sincos"); 167 setLibcallName(RTLIB::SINCOS_F80, "sincosl"); 168 setLibcallName(RTLIB::SINCOS_F128, "sincosl"); 169 setLibcallName(RTLIB::SINCOS_PPCF128, "sincosl"); 170 } 171 172 if (TT.isOSOpenBSD()) { 173 setLibcallName(RTLIB::STACKPROTECTOR_CHECK_FAIL, nullptr); 174 } 175 } 176 177 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 178 /// UNKNOWN_LIBCALL if there is none. 179 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 180 if (OpVT == MVT::f16) { 181 if (RetVT == MVT::f32) 182 return FPEXT_F16_F32; 183 } else if (OpVT == MVT::f32) { 184 if (RetVT == MVT::f64) 185 return FPEXT_F32_F64; 186 if (RetVT == MVT::f128) 187 return FPEXT_F32_F128; 188 if (RetVT == MVT::ppcf128) 189 return FPEXT_F32_PPCF128; 190 } else if (OpVT == MVT::f64) { 191 if (RetVT == MVT::f128) 192 return FPEXT_F64_F128; 193 else if (RetVT == MVT::ppcf128) 194 return FPEXT_F64_PPCF128; 195 } else if (OpVT == MVT::f80) { 196 if (RetVT == MVT::f128) 197 return FPEXT_F80_F128; 198 } 199 200 return UNKNOWN_LIBCALL; 201 } 202 203 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 204 /// UNKNOWN_LIBCALL if there is none. 205 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 206 if (RetVT == MVT::f16) { 207 if (OpVT == MVT::f32) 208 return FPROUND_F32_F16; 209 if (OpVT == MVT::f64) 210 return FPROUND_F64_F16; 211 if (OpVT == MVT::f80) 212 return FPROUND_F80_F16; 213 if (OpVT == MVT::f128) 214 return FPROUND_F128_F16; 215 if (OpVT == MVT::ppcf128) 216 return FPROUND_PPCF128_F16; 217 } else if (RetVT == MVT::f32) { 218 if (OpVT == MVT::f64) 219 return FPROUND_F64_F32; 220 if (OpVT == MVT::f80) 221 return FPROUND_F80_F32; 222 if (OpVT == MVT::f128) 223 return FPROUND_F128_F32; 224 if (OpVT == MVT::ppcf128) 225 return FPROUND_PPCF128_F32; 226 } else if (RetVT == MVT::f64) { 227 if (OpVT == MVT::f80) 228 return FPROUND_F80_F64; 229 if (OpVT == MVT::f128) 230 return FPROUND_F128_F64; 231 if (OpVT == MVT::ppcf128) 232 return FPROUND_PPCF128_F64; 233 } else if (RetVT == MVT::f80) { 234 if (OpVT == MVT::f128) 235 return FPROUND_F128_F80; 236 } 237 238 return UNKNOWN_LIBCALL; 239 } 240 241 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 242 /// UNKNOWN_LIBCALL if there is none. 243 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 244 if (OpVT == MVT::f32) { 245 if (RetVT == MVT::i32) 246 return FPTOSINT_F32_I32; 247 if (RetVT == MVT::i64) 248 return FPTOSINT_F32_I64; 249 if (RetVT == MVT::i128) 250 return FPTOSINT_F32_I128; 251 } else if (OpVT == MVT::f64) { 252 if (RetVT == MVT::i32) 253 return FPTOSINT_F64_I32; 254 if (RetVT == MVT::i64) 255 return FPTOSINT_F64_I64; 256 if (RetVT == MVT::i128) 257 return FPTOSINT_F64_I128; 258 } else if (OpVT == MVT::f80) { 259 if (RetVT == MVT::i32) 260 return FPTOSINT_F80_I32; 261 if (RetVT == MVT::i64) 262 return FPTOSINT_F80_I64; 263 if (RetVT == MVT::i128) 264 return FPTOSINT_F80_I128; 265 } else if (OpVT == MVT::f128) { 266 if (RetVT == MVT::i32) 267 return FPTOSINT_F128_I32; 268 if (RetVT == MVT::i64) 269 return FPTOSINT_F128_I64; 270 if (RetVT == MVT::i128) 271 return FPTOSINT_F128_I128; 272 } else if (OpVT == MVT::ppcf128) { 273 if (RetVT == MVT::i32) 274 return FPTOSINT_PPCF128_I32; 275 if (RetVT == MVT::i64) 276 return FPTOSINT_PPCF128_I64; 277 if (RetVT == MVT::i128) 278 return FPTOSINT_PPCF128_I128; 279 } 280 return UNKNOWN_LIBCALL; 281 } 282 283 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 284 /// UNKNOWN_LIBCALL if there is none. 285 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 286 if (OpVT == MVT::f32) { 287 if (RetVT == MVT::i32) 288 return FPTOUINT_F32_I32; 289 if (RetVT == MVT::i64) 290 return FPTOUINT_F32_I64; 291 if (RetVT == MVT::i128) 292 return FPTOUINT_F32_I128; 293 } else if (OpVT == MVT::f64) { 294 if (RetVT == MVT::i32) 295 return FPTOUINT_F64_I32; 296 if (RetVT == MVT::i64) 297 return FPTOUINT_F64_I64; 298 if (RetVT == MVT::i128) 299 return FPTOUINT_F64_I128; 300 } else if (OpVT == MVT::f80) { 301 if (RetVT == MVT::i32) 302 return FPTOUINT_F80_I32; 303 if (RetVT == MVT::i64) 304 return FPTOUINT_F80_I64; 305 if (RetVT == MVT::i128) 306 return FPTOUINT_F80_I128; 307 } else if (OpVT == MVT::f128) { 308 if (RetVT == MVT::i32) 309 return FPTOUINT_F128_I32; 310 if (RetVT == MVT::i64) 311 return FPTOUINT_F128_I64; 312 if (RetVT == MVT::i128) 313 return FPTOUINT_F128_I128; 314 } else if (OpVT == MVT::ppcf128) { 315 if (RetVT == MVT::i32) 316 return FPTOUINT_PPCF128_I32; 317 if (RetVT == MVT::i64) 318 return FPTOUINT_PPCF128_I64; 319 if (RetVT == MVT::i128) 320 return FPTOUINT_PPCF128_I128; 321 } 322 return UNKNOWN_LIBCALL; 323 } 324 325 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 326 /// UNKNOWN_LIBCALL if there is none. 327 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 328 if (OpVT == MVT::i32) { 329 if (RetVT == MVT::f32) 330 return SINTTOFP_I32_F32; 331 if (RetVT == MVT::f64) 332 return SINTTOFP_I32_F64; 333 if (RetVT == MVT::f80) 334 return SINTTOFP_I32_F80; 335 if (RetVT == MVT::f128) 336 return SINTTOFP_I32_F128; 337 if (RetVT == MVT::ppcf128) 338 return SINTTOFP_I32_PPCF128; 339 } else if (OpVT == MVT::i64) { 340 if (RetVT == MVT::f32) 341 return SINTTOFP_I64_F32; 342 if (RetVT == MVT::f64) 343 return SINTTOFP_I64_F64; 344 if (RetVT == MVT::f80) 345 return SINTTOFP_I64_F80; 346 if (RetVT == MVT::f128) 347 return SINTTOFP_I64_F128; 348 if (RetVT == MVT::ppcf128) 349 return SINTTOFP_I64_PPCF128; 350 } else if (OpVT == MVT::i128) { 351 if (RetVT == MVT::f32) 352 return SINTTOFP_I128_F32; 353 if (RetVT == MVT::f64) 354 return SINTTOFP_I128_F64; 355 if (RetVT == MVT::f80) 356 return SINTTOFP_I128_F80; 357 if (RetVT == MVT::f128) 358 return SINTTOFP_I128_F128; 359 if (RetVT == MVT::ppcf128) 360 return SINTTOFP_I128_PPCF128; 361 } 362 return UNKNOWN_LIBCALL; 363 } 364 365 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 366 /// UNKNOWN_LIBCALL if there is none. 367 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 368 if (OpVT == MVT::i32) { 369 if (RetVT == MVT::f32) 370 return UINTTOFP_I32_F32; 371 if (RetVT == MVT::f64) 372 return UINTTOFP_I32_F64; 373 if (RetVT == MVT::f80) 374 return UINTTOFP_I32_F80; 375 if (RetVT == MVT::f128) 376 return UINTTOFP_I32_F128; 377 if (RetVT == MVT::ppcf128) 378 return UINTTOFP_I32_PPCF128; 379 } else if (OpVT == MVT::i64) { 380 if (RetVT == MVT::f32) 381 return UINTTOFP_I64_F32; 382 if (RetVT == MVT::f64) 383 return UINTTOFP_I64_F64; 384 if (RetVT == MVT::f80) 385 return UINTTOFP_I64_F80; 386 if (RetVT == MVT::f128) 387 return UINTTOFP_I64_F128; 388 if (RetVT == MVT::ppcf128) 389 return UINTTOFP_I64_PPCF128; 390 } else if (OpVT == MVT::i128) { 391 if (RetVT == MVT::f32) 392 return UINTTOFP_I128_F32; 393 if (RetVT == MVT::f64) 394 return UINTTOFP_I128_F64; 395 if (RetVT == MVT::f80) 396 return UINTTOFP_I128_F80; 397 if (RetVT == MVT::f128) 398 return UINTTOFP_I128_F128; 399 if (RetVT == MVT::ppcf128) 400 return UINTTOFP_I128_PPCF128; 401 } 402 return UNKNOWN_LIBCALL; 403 } 404 405 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 406 #define OP_TO_LIBCALL(Name, Enum) \ 407 case Name: \ 408 switch (VT.SimpleTy) { \ 409 default: \ 410 return UNKNOWN_LIBCALL; \ 411 case MVT::i8: \ 412 return Enum##_1; \ 413 case MVT::i16: \ 414 return Enum##_2; \ 415 case MVT::i32: \ 416 return Enum##_4; \ 417 case MVT::i64: \ 418 return Enum##_8; \ 419 case MVT::i128: \ 420 return Enum##_16; \ 421 } 422 423 switch (Opc) { 424 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 425 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 426 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 427 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 428 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 429 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 430 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 431 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 432 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 433 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 434 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 435 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 436 } 437 438 #undef OP_TO_LIBCALL 439 440 return UNKNOWN_LIBCALL; 441 } 442 443 RTLIB::Libcall RTLIB::getMEMCPY_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 444 switch (ElementSize) { 445 case 1: 446 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_1; 447 case 2: 448 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_2; 449 case 4: 450 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_4; 451 case 8: 452 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_8; 453 case 16: 454 return MEMCPY_ELEMENT_UNORDERED_ATOMIC_16; 455 default: 456 return UNKNOWN_LIBCALL; 457 } 458 } 459 460 RTLIB::Libcall RTLIB::getMEMMOVE_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 461 switch (ElementSize) { 462 case 1: 463 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_1; 464 case 2: 465 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_2; 466 case 4: 467 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_4; 468 case 8: 469 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_8; 470 case 16: 471 return MEMMOVE_ELEMENT_UNORDERED_ATOMIC_16; 472 default: 473 return UNKNOWN_LIBCALL; 474 } 475 } 476 477 RTLIB::Libcall RTLIB::getMEMSET_ELEMENT_UNORDERED_ATOMIC(uint64_t ElementSize) { 478 switch (ElementSize) { 479 case 1: 480 return MEMSET_ELEMENT_UNORDERED_ATOMIC_1; 481 case 2: 482 return MEMSET_ELEMENT_UNORDERED_ATOMIC_2; 483 case 4: 484 return MEMSET_ELEMENT_UNORDERED_ATOMIC_4; 485 case 8: 486 return MEMSET_ELEMENT_UNORDERED_ATOMIC_8; 487 case 16: 488 return MEMSET_ELEMENT_UNORDERED_ATOMIC_16; 489 default: 490 return UNKNOWN_LIBCALL; 491 } 492 } 493 494 /// InitCmpLibcallCCs - Set default comparison libcall CC. 495 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 496 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 497 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 498 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 499 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 500 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 501 CCs[RTLIB::UNE_F32] = ISD::SETNE; 502 CCs[RTLIB::UNE_F64] = ISD::SETNE; 503 CCs[RTLIB::UNE_F128] = ISD::SETNE; 504 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 505 CCs[RTLIB::OGE_F32] = ISD::SETGE; 506 CCs[RTLIB::OGE_F64] = ISD::SETGE; 507 CCs[RTLIB::OGE_F128] = ISD::SETGE; 508 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 509 CCs[RTLIB::OLT_F32] = ISD::SETLT; 510 CCs[RTLIB::OLT_F64] = ISD::SETLT; 511 CCs[RTLIB::OLT_F128] = ISD::SETLT; 512 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 513 CCs[RTLIB::OLE_F32] = ISD::SETLE; 514 CCs[RTLIB::OLE_F64] = ISD::SETLE; 515 CCs[RTLIB::OLE_F128] = ISD::SETLE; 516 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 517 CCs[RTLIB::OGT_F32] = ISD::SETGT; 518 CCs[RTLIB::OGT_F64] = ISD::SETGT; 519 CCs[RTLIB::OGT_F128] = ISD::SETGT; 520 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 521 CCs[RTLIB::UO_F32] = ISD::SETNE; 522 CCs[RTLIB::UO_F64] = ISD::SETNE; 523 CCs[RTLIB::UO_F128] = ISD::SETNE; 524 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 525 CCs[RTLIB::O_F32] = ISD::SETEQ; 526 CCs[RTLIB::O_F64] = ISD::SETEQ; 527 CCs[RTLIB::O_F128] = ISD::SETEQ; 528 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 529 } 530 531 /// NOTE: The TargetMachine owns TLOF. 532 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 533 initActions(); 534 535 // Perform these initializations only once. 536 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 537 MaxLoadsPerMemcmp = 8; 538 MaxGluedStoresPerMemcpy = 0; 539 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize = 540 MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4; 541 UseUnderscoreSetJmp = false; 542 UseUnderscoreLongJmp = false; 543 HasMultipleConditionRegisters = false; 544 HasExtractBitsInsn = false; 545 JumpIsExpensive = JumpIsExpensiveOverride; 546 PredictableSelectIsExpensive = false; 547 EnableExtLdPromotion = false; 548 HasFloatingPointExceptions = true; 549 StackPointerRegisterToSaveRestore = 0; 550 BooleanContents = UndefinedBooleanContent; 551 BooleanFloatContents = UndefinedBooleanContent; 552 BooleanVectorContents = UndefinedBooleanContent; 553 SchedPreferenceInfo = Sched::ILP; 554 JumpBufSize = 0; 555 JumpBufAlignment = 0; 556 MinFunctionAlignment = 0; 557 PrefFunctionAlignment = 0; 558 PrefLoopAlignment = 0; 559 GatherAllAliasesMaxDepth = 18; 560 MinStackArgumentAlignment = 1; 561 // TODO: the default will be switched to 0 in the next commit, along 562 // with the Target-specific changes necessary. 563 MaxAtomicSizeInBitsSupported = 1024; 564 565 MinCmpXchgSizeInBits = 0; 566 SupportsUnalignedAtomics = false; 567 568 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 569 570 InitLibcalls(TM.getTargetTriple()); 571 InitCmpLibcallCCs(CmpLibcallCCs); 572 } 573 574 void TargetLoweringBase::initActions() { 575 // All operations default to being supported. 576 memset(OpActions, 0, sizeof(OpActions)); 577 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 578 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 579 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 580 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 581 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 582 std::fill(std::begin(TargetDAGCombineArray), 583 std::end(TargetDAGCombineArray), 0); 584 585 for (MVT VT : MVT::fp_valuetypes()) { 586 MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits()); 587 if (IntVT.isValid()) { 588 setOperationAction(ISD::ATOMIC_SWAP, VT, Promote); 589 AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT); 590 } 591 } 592 593 // Set default actions for various operations. 594 for (MVT VT : MVT::all_valuetypes()) { 595 // Default all indexed load / store to expand. 596 for (unsigned IM = (unsigned)ISD::PRE_INC; 597 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 598 setIndexedLoadAction(IM, VT, Expand); 599 setIndexedStoreAction(IM, VT, Expand); 600 } 601 602 // Most backends expect to see the node which just returns the value loaded. 603 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 604 605 // These operations default to expand. 606 setOperationAction(ISD::FGETSIGN, VT, Expand); 607 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 608 setOperationAction(ISD::FMINNUM, VT, Expand); 609 setOperationAction(ISD::FMAXNUM, VT, Expand); 610 setOperationAction(ISD::FMINNUM_IEEE, VT, Expand); 611 setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand); 612 setOperationAction(ISD::FMINIMUM, VT, Expand); 613 setOperationAction(ISD::FMAXIMUM, VT, Expand); 614 setOperationAction(ISD::FMAD, VT, Expand); 615 setOperationAction(ISD::SMIN, VT, Expand); 616 setOperationAction(ISD::SMAX, VT, Expand); 617 setOperationAction(ISD::UMIN, VT, Expand); 618 setOperationAction(ISD::UMAX, VT, Expand); 619 setOperationAction(ISD::ABS, VT, Expand); 620 setOperationAction(ISD::FSHL, VT, Expand); 621 setOperationAction(ISD::FSHR, VT, Expand); 622 setOperationAction(ISD::SADDSAT, VT, Expand); 623 setOperationAction(ISD::UADDSAT, VT, Expand); 624 setOperationAction(ISD::SSUBSAT, VT, Expand); 625 setOperationAction(ISD::USUBSAT, VT, Expand); 626 setOperationAction(ISD::SMULFIX, VT, Expand); 627 setOperationAction(ISD::UMULFIX, VT, Expand); 628 629 // Overflow operations default to expand 630 setOperationAction(ISD::SADDO, VT, Expand); 631 setOperationAction(ISD::SSUBO, VT, Expand); 632 setOperationAction(ISD::UADDO, VT, Expand); 633 setOperationAction(ISD::USUBO, VT, Expand); 634 setOperationAction(ISD::SMULO, VT, Expand); 635 setOperationAction(ISD::UMULO, VT, Expand); 636 637 // ADDCARRY operations default to expand 638 setOperationAction(ISD::ADDCARRY, VT, Expand); 639 setOperationAction(ISD::SUBCARRY, VT, Expand); 640 setOperationAction(ISD::SETCCCARRY, VT, Expand); 641 642 // ADDC/ADDE/SUBC/SUBE default to expand. 643 setOperationAction(ISD::ADDC, VT, Expand); 644 setOperationAction(ISD::ADDE, VT, Expand); 645 setOperationAction(ISD::SUBC, VT, Expand); 646 setOperationAction(ISD::SUBE, VT, Expand); 647 648 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 649 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 650 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 651 652 setOperationAction(ISD::BITREVERSE, VT, Expand); 653 654 // These library functions default to expand. 655 setOperationAction(ISD::FROUND, VT, Expand); 656 setOperationAction(ISD::FPOWI, VT, Expand); 657 658 // These operations default to expand for vector types. 659 if (VT.isVector()) { 660 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 661 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 662 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 663 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 664 } 665 666 // For most targets @llvm.get.dynamic.area.offset just returns 0. 667 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 668 } 669 670 // Most targets ignore the @llvm.prefetch intrinsic. 671 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 672 673 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 674 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 675 676 // ConstantFP nodes default to expand. Targets can either change this to 677 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 678 // to optimize expansions for certain constants. 679 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 680 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 681 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 682 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 683 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 684 685 // These library functions default to expand. 686 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 687 setOperationAction(ISD::FCBRT, VT, Expand); 688 setOperationAction(ISD::FLOG , VT, Expand); 689 setOperationAction(ISD::FLOG2, VT, Expand); 690 setOperationAction(ISD::FLOG10, VT, Expand); 691 setOperationAction(ISD::FEXP , VT, Expand); 692 setOperationAction(ISD::FEXP2, VT, Expand); 693 setOperationAction(ISD::FFLOOR, VT, Expand); 694 setOperationAction(ISD::FNEARBYINT, VT, Expand); 695 setOperationAction(ISD::FCEIL, VT, Expand); 696 setOperationAction(ISD::FRINT, VT, Expand); 697 setOperationAction(ISD::FTRUNC, VT, Expand); 698 setOperationAction(ISD::FROUND, VT, Expand); 699 } 700 701 // Default ISD::TRAP to expand (which turns it into abort). 702 setOperationAction(ISD::TRAP, MVT::Other, Expand); 703 704 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 705 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 706 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 707 } 708 709 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 710 EVT) const { 711 return MVT::getIntegerVT(8 * DL.getPointerSize(0)); 712 } 713 714 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 715 bool LegalTypes) const { 716 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 717 if (LHSTy.isVector()) 718 return LHSTy; 719 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 720 : getPointerTy(DL); 721 } 722 723 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 724 assert(isTypeLegal(VT)); 725 switch (Op) { 726 default: 727 return false; 728 case ISD::SDIV: 729 case ISD::UDIV: 730 case ISD::SREM: 731 case ISD::UREM: 732 return true; 733 } 734 } 735 736 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 737 // If the command-line option was specified, ignore this request. 738 if (!JumpIsExpensiveOverride.getNumOccurrences()) 739 JumpIsExpensive = isExpensive; 740 } 741 742 TargetLoweringBase::LegalizeKind 743 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 744 // If this is a simple type, use the ComputeRegisterProp mechanism. 745 if (VT.isSimple()) { 746 MVT SVT = VT.getSimpleVT(); 747 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 748 MVT NVT = TransformToType[SVT.SimpleTy]; 749 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 750 751 assert((LA == TypeLegal || LA == TypeSoftenFloat || 752 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) && 753 "Promote may not follow Expand or Promote"); 754 755 if (LA == TypeSplitVector) 756 return LegalizeKind(LA, 757 EVT::getVectorVT(Context, SVT.getVectorElementType(), 758 SVT.getVectorNumElements() / 2)); 759 if (LA == TypeScalarizeVector) 760 return LegalizeKind(LA, SVT.getVectorElementType()); 761 return LegalizeKind(LA, NVT); 762 } 763 764 // Handle Extended Scalar Types. 765 if (!VT.isVector()) { 766 assert(VT.isInteger() && "Float types must be simple"); 767 unsigned BitSize = VT.getSizeInBits(); 768 // First promote to a power-of-two size, then expand if necessary. 769 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 770 EVT NVT = VT.getRoundIntegerType(Context); 771 assert(NVT != VT && "Unable to round integer VT"); 772 LegalizeKind NextStep = getTypeConversion(Context, NVT); 773 // Avoid multi-step promotion. 774 if (NextStep.first == TypePromoteInteger) 775 return NextStep; 776 // Return rounded integer type. 777 return LegalizeKind(TypePromoteInteger, NVT); 778 } 779 780 return LegalizeKind(TypeExpandInteger, 781 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 782 } 783 784 // Handle vector types. 785 unsigned NumElts = VT.getVectorNumElements(); 786 EVT EltVT = VT.getVectorElementType(); 787 788 // Vectors with only one element are always scalarized. 789 if (NumElts == 1) 790 return LegalizeKind(TypeScalarizeVector, EltVT); 791 792 // Try to widen vector elements until the element type is a power of two and 793 // promote it to a legal type later on, for example: 794 // <3 x i8> -> <4 x i8> -> <4 x i32> 795 if (EltVT.isInteger()) { 796 // Vectors with a number of elements that is not a power of two are always 797 // widened, for example <3 x i8> -> <4 x i8>. 798 if (!VT.isPow2VectorType()) { 799 NumElts = (unsigned)NextPowerOf2(NumElts); 800 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 801 return LegalizeKind(TypeWidenVector, NVT); 802 } 803 804 // Examine the element type. 805 LegalizeKind LK = getTypeConversion(Context, EltVT); 806 807 // If type is to be expanded, split the vector. 808 // <4 x i140> -> <2 x i140> 809 if (LK.first == TypeExpandInteger) 810 return LegalizeKind(TypeSplitVector, 811 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 812 813 // Promote the integer element types until a legal vector type is found 814 // or until the element integer type is too big. If a legal type was not 815 // found, fallback to the usual mechanism of widening/splitting the 816 // vector. 817 EVT OldEltVT = EltVT; 818 while (true) { 819 // Increase the bitwidth of the element to the next pow-of-two 820 // (which is greater than 8 bits). 821 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 822 .getRoundIntegerType(Context); 823 824 // Stop trying when getting a non-simple element type. 825 // Note that vector elements may be greater than legal vector element 826 // types. Example: X86 XMM registers hold 64bit element on 32bit 827 // systems. 828 if (!EltVT.isSimple()) 829 break; 830 831 // Build a new vector type and check if it is legal. 832 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 833 // Found a legal promoted vector type. 834 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 835 return LegalizeKind(TypePromoteInteger, 836 EVT::getVectorVT(Context, EltVT, NumElts)); 837 } 838 839 // Reset the type to the unexpanded type if we did not find a legal vector 840 // type with a promoted vector element type. 841 EltVT = OldEltVT; 842 } 843 844 // Try to widen the vector until a legal type is found. 845 // If there is no wider legal type, split the vector. 846 while (true) { 847 // Round up to the next power of 2. 848 NumElts = (unsigned)NextPowerOf2(NumElts); 849 850 // If there is no simple vector type with this many elements then there 851 // cannot be a larger legal vector type. Note that this assumes that 852 // there are no skipped intermediate vector types in the simple types. 853 if (!EltVT.isSimple()) 854 break; 855 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 856 if (LargerVector == MVT()) 857 break; 858 859 // If this type is legal then widen the vector. 860 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 861 return LegalizeKind(TypeWidenVector, LargerVector); 862 } 863 864 // Widen odd vectors to next power of two. 865 if (!VT.isPow2VectorType()) { 866 EVT NVT = VT.getPow2VectorType(Context); 867 return LegalizeKind(TypeWidenVector, NVT); 868 } 869 870 // Vectors with illegal element types are expanded. 871 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 872 return LegalizeKind(TypeSplitVector, NVT); 873 } 874 875 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 876 unsigned &NumIntermediates, 877 MVT &RegisterVT, 878 TargetLoweringBase *TLI) { 879 // Figure out the right, legal destination reg to copy into. 880 unsigned NumElts = VT.getVectorNumElements(); 881 MVT EltTy = VT.getVectorElementType(); 882 883 unsigned NumVectorRegs = 1; 884 885 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 886 // could break down into LHS/RHS like LegalizeDAG does. 887 if (!isPowerOf2_32(NumElts)) { 888 NumVectorRegs = NumElts; 889 NumElts = 1; 890 } 891 892 // Divide the input until we get to a supported size. This will always 893 // end with a scalar if the target doesn't support vectors. 894 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 895 NumElts >>= 1; 896 NumVectorRegs <<= 1; 897 } 898 899 NumIntermediates = NumVectorRegs; 900 901 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 902 if (!TLI->isTypeLegal(NewVT)) 903 NewVT = EltTy; 904 IntermediateVT = NewVT; 905 906 unsigned NewVTSize = NewVT.getSizeInBits(); 907 908 // Convert sizes such as i33 to i64. 909 if (!isPowerOf2_32(NewVTSize)) 910 NewVTSize = NextPowerOf2(NewVTSize); 911 912 MVT DestVT = TLI->getRegisterType(NewVT); 913 RegisterVT = DestVT; 914 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 915 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 916 917 // Otherwise, promotion or legal types use the same number of registers as 918 // the vector decimated to the appropriate level. 919 return NumVectorRegs; 920 } 921 922 /// isLegalRC - Return true if the value types that can be represented by the 923 /// specified register class are all legal. 924 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 925 const TargetRegisterClass &RC) const { 926 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 927 if (isTypeLegal(*I)) 928 return true; 929 return false; 930 } 931 932 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 933 /// sequence of memory operands that is recognized by PrologEpilogInserter. 934 MachineBasicBlock * 935 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 936 MachineBasicBlock *MBB) const { 937 MachineInstr *MI = &InitialMI; 938 MachineFunction &MF = *MI->getMF(); 939 MachineFrameInfo &MFI = MF.getFrameInfo(); 940 941 // We're handling multiple types of operands here: 942 // PATCHPOINT MetaArgs - live-in, read only, direct 943 // STATEPOINT Deopt Spill - live-through, read only, indirect 944 // STATEPOINT Deopt Alloca - live-through, read only, direct 945 // (We're currently conservative and mark the deopt slots read/write in 946 // practice.) 947 // STATEPOINT GC Spill - live-through, read/write, indirect 948 // STATEPOINT GC Alloca - live-through, read/write, direct 949 // The live-in vs live-through is handled already (the live through ones are 950 // all stack slots), but we need to handle the different type of stackmap 951 // operands and memory effects here. 952 953 // MI changes inside this loop as we grow operands. 954 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 955 MachineOperand &MO = MI->getOperand(OperIdx); 956 if (!MO.isFI()) 957 continue; 958 959 // foldMemoryOperand builds a new MI after replacing a single FI operand 960 // with the canonical set of five x86 addressing-mode operands. 961 int FI = MO.getIndex(); 962 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 963 964 // Copy operands before the frame-index. 965 for (unsigned i = 0; i < OperIdx; ++i) 966 MIB.add(MI->getOperand(i)); 967 // Add frame index operands recognized by stackmaps.cpp 968 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 969 // indirect-mem-ref tag, size, #FI, offset. 970 // Used for spills inserted by StatepointLowering. This codepath is not 971 // used for patchpoints/stackmaps at all, for these spilling is done via 972 // foldMemoryOperand callback only. 973 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 974 MIB.addImm(StackMaps::IndirectMemRefOp); 975 MIB.addImm(MFI.getObjectSize(FI)); 976 MIB.add(MI->getOperand(OperIdx)); 977 MIB.addImm(0); 978 } else { 979 // direct-mem-ref tag, #FI, offset. 980 // Used by patchpoint, and direct alloca arguments to statepoints 981 MIB.addImm(StackMaps::DirectMemRefOp); 982 MIB.add(MI->getOperand(OperIdx)); 983 MIB.addImm(0); 984 } 985 // Copy the operands after the frame index. 986 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 987 MIB.add(MI->getOperand(i)); 988 989 // Inherit previous memory operands. 990 MIB.cloneMemRefs(*MI); 991 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 992 993 // Add a new memory operand for this FI. 994 assert(MFI.getObjectOffset(FI) != -1); 995 996 auto Flags = MachineMemOperand::MOLoad; 997 if (MI->getOpcode() == TargetOpcode::STATEPOINT) { 998 Flags |= MachineMemOperand::MOStore; 999 Flags |= MachineMemOperand::MOVolatile; 1000 } 1001 MachineMemOperand *MMO = MF.getMachineMemOperand( 1002 MachinePointerInfo::getFixedStack(MF, FI), Flags, 1003 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 1004 MIB->addMemOperand(MF, MMO); 1005 1006 // Replace the instruction and update the operand index. 1007 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1008 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1009 MI->eraseFromParent(); 1010 MI = MIB; 1011 } 1012 return MBB; 1013 } 1014 1015 MachineBasicBlock * 1016 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1017 MachineBasicBlock *MBB) const { 1018 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1019 "Called emitXRayCustomEvent on the wrong MI!"); 1020 auto &MF = *MI.getMF(); 1021 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1022 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1023 MIB.add(MI.getOperand(OpIdx)); 1024 1025 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1026 MI.eraseFromParent(); 1027 return MBB; 1028 } 1029 1030 MachineBasicBlock * 1031 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1032 MachineBasicBlock *MBB) const { 1033 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1034 "Called emitXRayTypedEvent on the wrong MI!"); 1035 auto &MF = *MI.getMF(); 1036 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1037 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1038 MIB.add(MI.getOperand(OpIdx)); 1039 1040 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1041 MI.eraseFromParent(); 1042 return MBB; 1043 } 1044 1045 /// findRepresentativeClass - Return the largest legal super-reg register class 1046 /// of the register class for the specified type and its associated "cost". 1047 // This function is in TargetLowering because it uses RegClassForVT which would 1048 // need to be moved to TargetRegisterInfo and would necessitate moving 1049 // isTypeLegal over as well - a massive change that would just require 1050 // TargetLowering having a TargetRegisterInfo class member that it would use. 1051 std::pair<const TargetRegisterClass *, uint8_t> 1052 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1053 MVT VT) const { 1054 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1055 if (!RC) 1056 return std::make_pair(RC, 0); 1057 1058 // Compute the set of all super-register classes. 1059 BitVector SuperRegRC(TRI->getNumRegClasses()); 1060 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1061 SuperRegRC.setBitsInMask(RCI.getMask()); 1062 1063 // Find the first legal register class with the largest spill size. 1064 const TargetRegisterClass *BestRC = RC; 1065 for (unsigned i : SuperRegRC.set_bits()) { 1066 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1067 // We want the largest possible spill size. 1068 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1069 continue; 1070 if (!isLegalRC(*TRI, *SuperRC)) 1071 continue; 1072 BestRC = SuperRC; 1073 } 1074 return std::make_pair(BestRC, 1); 1075 } 1076 1077 /// computeRegisterProperties - Once all of the register classes are added, 1078 /// this allows us to compute derived properties we expose. 1079 void TargetLoweringBase::computeRegisterProperties( 1080 const TargetRegisterInfo *TRI) { 1081 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1082 "Too many value types for ValueTypeActions to hold!"); 1083 1084 // Everything defaults to needing one register. 1085 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1086 NumRegistersForVT[i] = 1; 1087 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1088 } 1089 // ...except isVoid, which doesn't need any registers. 1090 NumRegistersForVT[MVT::isVoid] = 0; 1091 1092 // Find the largest integer register class. 1093 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1094 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1095 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1096 1097 // Every integer value type larger than this largest register takes twice as 1098 // many registers to represent as the previous ValueType. 1099 for (unsigned ExpandedReg = LargestIntReg + 1; 1100 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1101 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1102 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1103 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1104 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1105 TypeExpandInteger); 1106 } 1107 1108 // Inspect all of the ValueType's smaller than the largest integer 1109 // register to see which ones need promotion. 1110 unsigned LegalIntReg = LargestIntReg; 1111 for (unsigned IntReg = LargestIntReg - 1; 1112 IntReg >= (unsigned)MVT::i1; --IntReg) { 1113 MVT IVT = (MVT::SimpleValueType)IntReg; 1114 if (isTypeLegal(IVT)) { 1115 LegalIntReg = IntReg; 1116 } else { 1117 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1118 (MVT::SimpleValueType)LegalIntReg; 1119 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1120 } 1121 } 1122 1123 // ppcf128 type is really two f64's. 1124 if (!isTypeLegal(MVT::ppcf128)) { 1125 if (isTypeLegal(MVT::f64)) { 1126 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1127 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1128 TransformToType[MVT::ppcf128] = MVT::f64; 1129 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1130 } else { 1131 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1132 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1133 TransformToType[MVT::ppcf128] = MVT::i128; 1134 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1135 } 1136 } 1137 1138 // Decide how to handle f128. If the target does not have native f128 support, 1139 // expand it to i128 and we will be generating soft float library calls. 1140 if (!isTypeLegal(MVT::f128)) { 1141 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1142 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1143 TransformToType[MVT::f128] = MVT::i128; 1144 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1145 } 1146 1147 // Decide how to handle f64. If the target does not have native f64 support, 1148 // expand it to i64 and we will be generating soft float library calls. 1149 if (!isTypeLegal(MVT::f64)) { 1150 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1151 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1152 TransformToType[MVT::f64] = MVT::i64; 1153 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1154 } 1155 1156 // Decide how to handle f32. If the target does not have native f32 support, 1157 // expand it to i32 and we will be generating soft float library calls. 1158 if (!isTypeLegal(MVT::f32)) { 1159 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1160 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1161 TransformToType[MVT::f32] = MVT::i32; 1162 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1163 } 1164 1165 // Decide how to handle f16. If the target does not have native f16 support, 1166 // promote it to f32, because there are no f16 library calls (except for 1167 // conversions). 1168 if (!isTypeLegal(MVT::f16)) { 1169 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1170 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1171 TransformToType[MVT::f16] = MVT::f32; 1172 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1173 } 1174 1175 // Loop over all of the vector value types to see which need transformations. 1176 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1177 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1178 MVT VT = (MVT::SimpleValueType) i; 1179 if (isTypeLegal(VT)) 1180 continue; 1181 1182 MVT EltVT = VT.getVectorElementType(); 1183 unsigned NElts = VT.getVectorNumElements(); 1184 bool IsLegalWiderType = false; 1185 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1186 switch (PreferredAction) { 1187 case TypePromoteInteger: 1188 // Try to promote the elements of integer vectors. If no legal 1189 // promotion was found, fall through to the widen-vector method. 1190 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1191 MVT SVT = (MVT::SimpleValueType) nVT; 1192 // Promote vectors of integers to vectors with the same number 1193 // of elements, with a wider element type. 1194 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1195 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1196 TransformToType[i] = SVT; 1197 RegisterTypeForVT[i] = SVT; 1198 NumRegistersForVT[i] = 1; 1199 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1200 IsLegalWiderType = true; 1201 break; 1202 } 1203 } 1204 if (IsLegalWiderType) 1205 break; 1206 LLVM_FALLTHROUGH; 1207 1208 case TypeWidenVector: 1209 // Try to widen the vector. 1210 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1211 MVT SVT = (MVT::SimpleValueType) nVT; 1212 if (SVT.getVectorElementType() == EltVT 1213 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1214 TransformToType[i] = SVT; 1215 RegisterTypeForVT[i] = SVT; 1216 NumRegistersForVT[i] = 1; 1217 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1218 IsLegalWiderType = true; 1219 break; 1220 } 1221 } 1222 if (IsLegalWiderType) 1223 break; 1224 LLVM_FALLTHROUGH; 1225 1226 case TypeSplitVector: 1227 case TypeScalarizeVector: { 1228 MVT IntermediateVT; 1229 MVT RegisterVT; 1230 unsigned NumIntermediates; 1231 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1232 NumIntermediates, RegisterVT, this); 1233 RegisterTypeForVT[i] = RegisterVT; 1234 1235 MVT NVT = VT.getPow2VectorType(); 1236 if (NVT == VT) { 1237 // Type is already a power of 2. The default action is to split. 1238 TransformToType[i] = MVT::Other; 1239 if (PreferredAction == TypeScalarizeVector) 1240 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1241 else if (PreferredAction == TypeSplitVector) 1242 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1243 else 1244 // Set type action according to the number of elements. 1245 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1246 : TypeSplitVector); 1247 } else { 1248 TransformToType[i] = NVT; 1249 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1250 } 1251 break; 1252 } 1253 default: 1254 llvm_unreachable("Unknown vector legalization action!"); 1255 } 1256 } 1257 1258 // Determine the 'representative' register class for each value type. 1259 // An representative register class is the largest (meaning one which is 1260 // not a sub-register class / subreg register class) legal register class for 1261 // a group of value types. For example, on i386, i8, i16, and i32 1262 // representative would be GR32; while on x86_64 it's GR64. 1263 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1264 const TargetRegisterClass* RRC; 1265 uint8_t Cost; 1266 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1267 RepRegClassForVT[i] = RRC; 1268 RepRegClassCostForVT[i] = Cost; 1269 } 1270 } 1271 1272 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1273 EVT VT) const { 1274 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1275 return getPointerTy(DL).SimpleTy; 1276 } 1277 1278 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1279 return MVT::i32; // return the default value 1280 } 1281 1282 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1283 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1284 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1285 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1286 /// 1287 /// This method returns the number of registers needed, and the VT for each 1288 /// register. It also returns the VT and quantity of the intermediate values 1289 /// before they are promoted/expanded. 1290 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1291 EVT &IntermediateVT, 1292 unsigned &NumIntermediates, 1293 MVT &RegisterVT) const { 1294 unsigned NumElts = VT.getVectorNumElements(); 1295 1296 // If there is a wider vector type with the same element type as this one, 1297 // or a promoted vector type that has the same number of elements which 1298 // are wider, then we should convert to that legal vector type. 1299 // This handles things like <2 x float> -> <4 x float> and 1300 // <4 x i1> -> <4 x i32>. 1301 LegalizeTypeAction TA = getTypeAction(Context, VT); 1302 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1303 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1304 if (isTypeLegal(RegisterEVT)) { 1305 IntermediateVT = RegisterEVT; 1306 RegisterVT = RegisterEVT.getSimpleVT(); 1307 NumIntermediates = 1; 1308 return 1; 1309 } 1310 } 1311 1312 // Figure out the right, legal destination reg to copy into. 1313 EVT EltTy = VT.getVectorElementType(); 1314 1315 unsigned NumVectorRegs = 1; 1316 1317 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1318 // could break down into LHS/RHS like LegalizeDAG does. 1319 if (!isPowerOf2_32(NumElts)) { 1320 NumVectorRegs = NumElts; 1321 NumElts = 1; 1322 } 1323 1324 // Divide the input until we get to a supported size. This will always 1325 // end with a scalar if the target doesn't support vectors. 1326 while (NumElts > 1 && !isTypeLegal( 1327 EVT::getVectorVT(Context, EltTy, NumElts))) { 1328 NumElts >>= 1; 1329 NumVectorRegs <<= 1; 1330 } 1331 1332 NumIntermediates = NumVectorRegs; 1333 1334 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1335 if (!isTypeLegal(NewVT)) 1336 NewVT = EltTy; 1337 IntermediateVT = NewVT; 1338 1339 MVT DestVT = getRegisterType(Context, NewVT); 1340 RegisterVT = DestVT; 1341 unsigned NewVTSize = NewVT.getSizeInBits(); 1342 1343 // Convert sizes such as i33 to i64. 1344 if (!isPowerOf2_32(NewVTSize)) 1345 NewVTSize = NextPowerOf2(NewVTSize); 1346 1347 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1348 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1349 1350 // Otherwise, promotion or legal types use the same number of registers as 1351 // the vector decimated to the appropriate level. 1352 return NumVectorRegs; 1353 } 1354 1355 /// Get the EVTs and ArgFlags collections that represent the legalized return 1356 /// type of the given function. This does not require a DAG or a return value, 1357 /// and is suitable for use before any DAGs for the function are constructed. 1358 /// TODO: Move this out of TargetLowering.cpp. 1359 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1360 AttributeList attr, 1361 SmallVectorImpl<ISD::OutputArg> &Outs, 1362 const TargetLowering &TLI, const DataLayout &DL) { 1363 SmallVector<EVT, 4> ValueVTs; 1364 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1365 unsigned NumValues = ValueVTs.size(); 1366 if (NumValues == 0) return; 1367 1368 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1369 EVT VT = ValueVTs[j]; 1370 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1371 1372 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1373 ExtendKind = ISD::SIGN_EXTEND; 1374 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1375 ExtendKind = ISD::ZERO_EXTEND; 1376 1377 // FIXME: C calling convention requires the return type to be promoted to 1378 // at least 32-bit. But this is not necessary for non-C calling 1379 // conventions. The frontend should mark functions whose return values 1380 // require promoting with signext or zeroext attributes. 1381 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1382 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1383 if (VT.bitsLT(MinVT)) 1384 VT = MinVT; 1385 } 1386 1387 unsigned NumParts = 1388 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1389 MVT PartVT = 1390 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1391 1392 // 'inreg' on function refers to return value 1393 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1394 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1395 Flags.setInReg(); 1396 1397 // Propagate extension type if any 1398 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1399 Flags.setSExt(); 1400 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1401 Flags.setZExt(); 1402 1403 for (unsigned i = 0; i < NumParts; ++i) 1404 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0)); 1405 } 1406 } 1407 1408 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1409 /// function arguments in the caller parameter area. This is the actual 1410 /// alignment, not its logarithm. 1411 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1412 const DataLayout &DL) const { 1413 return DL.getABITypeAlignment(Ty); 1414 } 1415 1416 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1417 const DataLayout &DL, EVT VT, 1418 unsigned AddrSpace, 1419 unsigned Alignment, 1420 bool *Fast) const { 1421 // Check if the specified alignment is sufficient based on the data layout. 1422 // TODO: While using the data layout works in practice, a better solution 1423 // would be to implement this check directly (make this a virtual function). 1424 // For example, the ABI alignment may change based on software platform while 1425 // this function should only be affected by hardware implementation. 1426 Type *Ty = VT.getTypeForEVT(Context); 1427 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1428 // Assume that an access that meets the ABI-specified alignment is fast. 1429 if (Fast != nullptr) 1430 *Fast = true; 1431 return true; 1432 } 1433 1434 // This is a misaligned access. 1435 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast); 1436 } 1437 1438 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1439 return BranchProbability(MinPercentageForPredictableBranch, 100); 1440 } 1441 1442 //===----------------------------------------------------------------------===// 1443 // TargetTransformInfo Helpers 1444 //===----------------------------------------------------------------------===// 1445 1446 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1447 enum InstructionOpcodes { 1448 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1449 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1450 #include "llvm/IR/Instruction.def" 1451 }; 1452 switch (static_cast<InstructionOpcodes>(Opcode)) { 1453 case Ret: return 0; 1454 case Br: return 0; 1455 case Switch: return 0; 1456 case IndirectBr: return 0; 1457 case Invoke: return 0; 1458 case CallBr: return 0; 1459 case Resume: return 0; 1460 case Unreachable: return 0; 1461 case CleanupRet: return 0; 1462 case CatchRet: return 0; 1463 case CatchPad: return 0; 1464 case CatchSwitch: return 0; 1465 case CleanupPad: return 0; 1466 case FNeg: return ISD::FNEG; 1467 case Add: return ISD::ADD; 1468 case FAdd: return ISD::FADD; 1469 case Sub: return ISD::SUB; 1470 case FSub: return ISD::FSUB; 1471 case Mul: return ISD::MUL; 1472 case FMul: return ISD::FMUL; 1473 case UDiv: return ISD::UDIV; 1474 case SDiv: return ISD::SDIV; 1475 case FDiv: return ISD::FDIV; 1476 case URem: return ISD::UREM; 1477 case SRem: return ISD::SREM; 1478 case FRem: return ISD::FREM; 1479 case Shl: return ISD::SHL; 1480 case LShr: return ISD::SRL; 1481 case AShr: return ISD::SRA; 1482 case And: return ISD::AND; 1483 case Or: return ISD::OR; 1484 case Xor: return ISD::XOR; 1485 case Alloca: return 0; 1486 case Load: return ISD::LOAD; 1487 case Store: return ISD::STORE; 1488 case GetElementPtr: return 0; 1489 case Fence: return 0; 1490 case AtomicCmpXchg: return 0; 1491 case AtomicRMW: return 0; 1492 case Trunc: return ISD::TRUNCATE; 1493 case ZExt: return ISD::ZERO_EXTEND; 1494 case SExt: return ISD::SIGN_EXTEND; 1495 case FPToUI: return ISD::FP_TO_UINT; 1496 case FPToSI: return ISD::FP_TO_SINT; 1497 case UIToFP: return ISD::UINT_TO_FP; 1498 case SIToFP: return ISD::SINT_TO_FP; 1499 case FPTrunc: return ISD::FP_ROUND; 1500 case FPExt: return ISD::FP_EXTEND; 1501 case PtrToInt: return ISD::BITCAST; 1502 case IntToPtr: return ISD::BITCAST; 1503 case BitCast: return ISD::BITCAST; 1504 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1505 case ICmp: return ISD::SETCC; 1506 case FCmp: return ISD::SETCC; 1507 case PHI: return 0; 1508 case Call: return 0; 1509 case Select: return ISD::SELECT; 1510 case UserOp1: return 0; 1511 case UserOp2: return 0; 1512 case VAArg: return 0; 1513 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1514 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1515 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1516 case ExtractValue: return ISD::MERGE_VALUES; 1517 case InsertValue: return ISD::MERGE_VALUES; 1518 case LandingPad: return 0; 1519 } 1520 1521 llvm_unreachable("Unknown instruction type encountered!"); 1522 } 1523 1524 std::pair<int, MVT> 1525 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1526 Type *Ty) const { 1527 LLVMContext &C = Ty->getContext(); 1528 EVT MTy = getValueType(DL, Ty); 1529 1530 int Cost = 1; 1531 // We keep legalizing the type until we find a legal kind. We assume that 1532 // the only operation that costs anything is the split. After splitting 1533 // we need to handle two types. 1534 while (true) { 1535 LegalizeKind LK = getTypeConversion(C, MTy); 1536 1537 if (LK.first == TypeLegal) 1538 return std::make_pair(Cost, MTy.getSimpleVT()); 1539 1540 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1541 Cost *= 2; 1542 1543 // Do not loop with f128 type. 1544 if (MTy == LK.second) 1545 return std::make_pair(Cost, MTy.getSimpleVT()); 1546 1547 // Keep legalizing the type. 1548 MTy = LK.second; 1549 } 1550 } 1551 1552 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1553 bool UseTLS) const { 1554 // compiler-rt provides a variable with a magic name. Targets that do not 1555 // link with compiler-rt may also provide such a variable. 1556 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1557 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1558 auto UnsafeStackPtr = 1559 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1560 1561 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1562 1563 if (!UnsafeStackPtr) { 1564 auto TLSModel = UseTLS ? 1565 GlobalValue::InitialExecTLSModel : 1566 GlobalValue::NotThreadLocal; 1567 // The global variable is not defined yet, define it ourselves. 1568 // We use the initial-exec TLS model because we do not support the 1569 // variable living anywhere other than in the main executable. 1570 UnsafeStackPtr = new GlobalVariable( 1571 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1572 UnsafeStackPtrVar, nullptr, TLSModel); 1573 } else { 1574 // The variable exists, check its type and attributes. 1575 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1576 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1577 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1578 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1579 (UseTLS ? "" : "not ") + "be thread-local"); 1580 } 1581 return UnsafeStackPtr; 1582 } 1583 1584 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1585 if (!TM.getTargetTriple().isAndroid()) 1586 return getDefaultSafeStackPointerLocation(IRB, true); 1587 1588 // Android provides a libc function to retrieve the address of the current 1589 // thread's unsafe stack pointer. 1590 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1591 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1592 FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address", 1593 StackPtrTy->getPointerTo(0)); 1594 return IRB.CreateCall(Fn); 1595 } 1596 1597 //===----------------------------------------------------------------------===// 1598 // Loop Strength Reduction hooks 1599 //===----------------------------------------------------------------------===// 1600 1601 /// isLegalAddressingMode - Return true if the addressing mode represented 1602 /// by AM is legal for this target, for a load/store of the specified type. 1603 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1604 const AddrMode &AM, Type *Ty, 1605 unsigned AS, Instruction *I) const { 1606 // The default implementation of this implements a conservative RISCy, r+r and 1607 // r+i addr mode. 1608 1609 // Allows a sign-extended 16-bit immediate field. 1610 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1611 return false; 1612 1613 // No global is ever allowed as a base. 1614 if (AM.BaseGV) 1615 return false; 1616 1617 // Only support r+r, 1618 switch (AM.Scale) { 1619 case 0: // "r+i" or just "i", depending on HasBaseReg. 1620 break; 1621 case 1: 1622 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1623 return false; 1624 // Otherwise we have r+r or r+i. 1625 break; 1626 case 2: 1627 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1628 return false; 1629 // Allow 2*r as r+r. 1630 break; 1631 default: // Don't allow n * r 1632 return false; 1633 } 1634 1635 return true; 1636 } 1637 1638 //===----------------------------------------------------------------------===// 1639 // Stack Protector 1640 //===----------------------------------------------------------------------===// 1641 1642 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1643 // so that SelectionDAG handle SSP. 1644 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1645 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1646 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1647 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1648 return M.getOrInsertGlobal("__guard_local", PtrTy); 1649 } 1650 return nullptr; 1651 } 1652 1653 // Currently only support "standard" __stack_chk_guard. 1654 // TODO: add LOAD_STACK_GUARD support. 1655 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1656 if (!M.getNamedValue("__stack_chk_guard")) 1657 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1658 GlobalVariable::ExternalLinkage, 1659 nullptr, "__stack_chk_guard"); 1660 } 1661 1662 // Currently only support "standard" __stack_chk_guard. 1663 // TODO: add LOAD_STACK_GUARD support. 1664 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1665 return M.getNamedValue("__stack_chk_guard"); 1666 } 1667 1668 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1669 return nullptr; 1670 } 1671 1672 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1673 return MinimumJumpTableEntries; 1674 } 1675 1676 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1677 MinimumJumpTableEntries = Val; 1678 } 1679 1680 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1681 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1682 } 1683 1684 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1685 return MaximumJumpTableSize; 1686 } 1687 1688 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1689 MaximumJumpTableSize = Val; 1690 } 1691 1692 //===----------------------------------------------------------------------===// 1693 // Reciprocal Estimates 1694 //===----------------------------------------------------------------------===// 1695 1696 /// Get the reciprocal estimate attribute string for a function that will 1697 /// override the target defaults. 1698 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1699 const Function &F = MF.getFunction(); 1700 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1701 } 1702 1703 /// Construct a string for the given reciprocal operation of the given type. 1704 /// This string should match the corresponding option to the front-end's 1705 /// "-mrecip" flag assuming those strings have been passed through in an 1706 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1707 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1708 std::string Name = VT.isVector() ? "vec-" : ""; 1709 1710 Name += IsSqrt ? "sqrt" : "div"; 1711 1712 // TODO: Handle "half" or other float types? 1713 if (VT.getScalarType() == MVT::f64) { 1714 Name += "d"; 1715 } else { 1716 assert(VT.getScalarType() == MVT::f32 && 1717 "Unexpected FP type for reciprocal estimate"); 1718 Name += "f"; 1719 } 1720 1721 return Name; 1722 } 1723 1724 /// Return the character position and value (a single numeric character) of a 1725 /// customized refinement operation in the input string if it exists. Return 1726 /// false if there is no customized refinement step count. 1727 static bool parseRefinementStep(StringRef In, size_t &Position, 1728 uint8_t &Value) { 1729 const char RefStepToken = ':'; 1730 Position = In.find(RefStepToken); 1731 if (Position == StringRef::npos) 1732 return false; 1733 1734 StringRef RefStepString = In.substr(Position + 1); 1735 // Allow exactly one numeric character for the additional refinement 1736 // step parameter. 1737 if (RefStepString.size() == 1) { 1738 char RefStepChar = RefStepString[0]; 1739 if (RefStepChar >= '0' && RefStepChar <= '9') { 1740 Value = RefStepChar - '0'; 1741 return true; 1742 } 1743 } 1744 report_fatal_error("Invalid refinement step for -recip."); 1745 } 1746 1747 /// For the input attribute string, return one of the ReciprocalEstimate enum 1748 /// status values (enabled, disabled, or not specified) for this operation on 1749 /// the specified data type. 1750 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1751 if (Override.empty()) 1752 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1753 1754 SmallVector<StringRef, 4> OverrideVector; 1755 Override.split(OverrideVector, ','); 1756 unsigned NumArgs = OverrideVector.size(); 1757 1758 // Check if "all", "none", or "default" was specified. 1759 if (NumArgs == 1) { 1760 // Look for an optional setting of the number of refinement steps needed 1761 // for this type of reciprocal operation. 1762 size_t RefPos; 1763 uint8_t RefSteps; 1764 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1765 // Split the string for further processing. 1766 Override = Override.substr(0, RefPos); 1767 } 1768 1769 // All reciprocal types are enabled. 1770 if (Override == "all") 1771 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1772 1773 // All reciprocal types are disabled. 1774 if (Override == "none") 1775 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1776 1777 // Target defaults for enablement are used. 1778 if (Override == "default") 1779 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1780 } 1781 1782 // The attribute string may omit the size suffix ('f'/'d'). 1783 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1784 std::string VTNameNoSize = VTName; 1785 VTNameNoSize.pop_back(); 1786 static const char DisabledPrefix = '!'; 1787 1788 for (StringRef RecipType : OverrideVector) { 1789 size_t RefPos; 1790 uint8_t RefSteps; 1791 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1792 RecipType = RecipType.substr(0, RefPos); 1793 1794 // Ignore the disablement token for string matching. 1795 bool IsDisabled = RecipType[0] == DisabledPrefix; 1796 if (IsDisabled) 1797 RecipType = RecipType.substr(1); 1798 1799 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1800 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1801 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1802 } 1803 1804 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1805 } 1806 1807 /// For the input attribute string, return the customized refinement step count 1808 /// for this operation on the specified data type. If the step count does not 1809 /// exist, return the ReciprocalEstimate enum value for unspecified. 1810 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1811 if (Override.empty()) 1812 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1813 1814 SmallVector<StringRef, 4> OverrideVector; 1815 Override.split(OverrideVector, ','); 1816 unsigned NumArgs = OverrideVector.size(); 1817 1818 // Check if "all", "default", or "none" was specified. 1819 if (NumArgs == 1) { 1820 // Look for an optional setting of the number of refinement steps needed 1821 // for this type of reciprocal operation. 1822 size_t RefPos; 1823 uint8_t RefSteps; 1824 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1825 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1826 1827 // Split the string for further processing. 1828 Override = Override.substr(0, RefPos); 1829 assert(Override != "none" && 1830 "Disabled reciprocals, but specifed refinement steps?"); 1831 1832 // If this is a general override, return the specified number of steps. 1833 if (Override == "all" || Override == "default") 1834 return RefSteps; 1835 } 1836 1837 // The attribute string may omit the size suffix ('f'/'d'). 1838 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1839 std::string VTNameNoSize = VTName; 1840 VTNameNoSize.pop_back(); 1841 1842 for (StringRef RecipType : OverrideVector) { 1843 size_t RefPos; 1844 uint8_t RefSteps; 1845 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1846 continue; 1847 1848 RecipType = RecipType.substr(0, RefPos); 1849 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1850 return RefSteps; 1851 } 1852 1853 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1854 } 1855 1856 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1857 MachineFunction &MF) const { 1858 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1859 } 1860 1861 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1862 MachineFunction &MF) const { 1863 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1864 } 1865 1866 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1867 MachineFunction &MF) const { 1868 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1869 } 1870 1871 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1872 MachineFunction &MF) const { 1873 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1874 } 1875 1876 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1877 MF.getRegInfo().freezeReservedRegs(MF); 1878 } 1879