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); 627 628 // Overflow operations default to expand 629 setOperationAction(ISD::SADDO, VT, Expand); 630 setOperationAction(ISD::SSUBO, VT, Expand); 631 setOperationAction(ISD::UADDO, VT, Expand); 632 setOperationAction(ISD::USUBO, VT, Expand); 633 setOperationAction(ISD::SMULO, VT, Expand); 634 setOperationAction(ISD::UMULO, VT, Expand); 635 636 // ADDCARRY operations default to expand 637 setOperationAction(ISD::ADDCARRY, VT, Expand); 638 setOperationAction(ISD::SUBCARRY, VT, Expand); 639 setOperationAction(ISD::SETCCCARRY, VT, Expand); 640 641 // ADDC/ADDE/SUBC/SUBE default to expand. 642 setOperationAction(ISD::ADDC, VT, Expand); 643 setOperationAction(ISD::ADDE, VT, Expand); 644 setOperationAction(ISD::SUBC, VT, Expand); 645 setOperationAction(ISD::SUBE, VT, Expand); 646 647 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 648 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 649 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 650 651 setOperationAction(ISD::BITREVERSE, VT, Expand); 652 653 // These library functions default to expand. 654 setOperationAction(ISD::FROUND, VT, Expand); 655 setOperationAction(ISD::FPOWI, VT, Expand); 656 657 // These operations default to expand for vector types. 658 if (VT.isVector()) { 659 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 660 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 661 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 662 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 663 } 664 665 // For most targets @llvm.get.dynamic.area.offset just returns 0. 666 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 667 } 668 669 // Most targets ignore the @llvm.prefetch intrinsic. 670 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 671 672 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 673 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 674 675 // ConstantFP nodes default to expand. Targets can either change this to 676 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 677 // to optimize expansions for certain constants. 678 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 679 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 680 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 681 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 682 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 683 684 // These library functions default to expand. 685 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 686 setOperationAction(ISD::FCBRT, VT, Expand); 687 setOperationAction(ISD::FLOG , VT, Expand); 688 setOperationAction(ISD::FLOG2, VT, Expand); 689 setOperationAction(ISD::FLOG10, VT, Expand); 690 setOperationAction(ISD::FEXP , VT, Expand); 691 setOperationAction(ISD::FEXP2, VT, Expand); 692 setOperationAction(ISD::FFLOOR, VT, Expand); 693 setOperationAction(ISD::FNEARBYINT, VT, Expand); 694 setOperationAction(ISD::FCEIL, VT, Expand); 695 setOperationAction(ISD::FRINT, VT, Expand); 696 setOperationAction(ISD::FTRUNC, VT, Expand); 697 setOperationAction(ISD::FROUND, VT, Expand); 698 } 699 700 // Default ISD::TRAP to expand (which turns it into abort). 701 setOperationAction(ISD::TRAP, MVT::Other, Expand); 702 703 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 704 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 705 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 706 } 707 708 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 709 EVT) const { 710 return MVT::getIntegerVT(8 * DL.getPointerSize(0)); 711 } 712 713 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 714 bool LegalTypes) const { 715 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 716 if (LHSTy.isVector()) 717 return LHSTy; 718 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 719 : getPointerTy(DL); 720 } 721 722 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 723 assert(isTypeLegal(VT)); 724 switch (Op) { 725 default: 726 return false; 727 case ISD::SDIV: 728 case ISD::UDIV: 729 case ISD::SREM: 730 case ISD::UREM: 731 return true; 732 } 733 } 734 735 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 736 // If the command-line option was specified, ignore this request. 737 if (!JumpIsExpensiveOverride.getNumOccurrences()) 738 JumpIsExpensive = isExpensive; 739 } 740 741 TargetLoweringBase::LegalizeKind 742 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 743 // If this is a simple type, use the ComputeRegisterProp mechanism. 744 if (VT.isSimple()) { 745 MVT SVT = VT.getSimpleVT(); 746 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 747 MVT NVT = TransformToType[SVT.SimpleTy]; 748 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 749 750 assert((LA == TypeLegal || LA == TypeSoftenFloat || 751 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) && 752 "Promote may not follow Expand or Promote"); 753 754 if (LA == TypeSplitVector) 755 return LegalizeKind(LA, 756 EVT::getVectorVT(Context, SVT.getVectorElementType(), 757 SVT.getVectorNumElements() / 2)); 758 if (LA == TypeScalarizeVector) 759 return LegalizeKind(LA, SVT.getVectorElementType()); 760 return LegalizeKind(LA, NVT); 761 } 762 763 // Handle Extended Scalar Types. 764 if (!VT.isVector()) { 765 assert(VT.isInteger() && "Float types must be simple"); 766 unsigned BitSize = VT.getSizeInBits(); 767 // First promote to a power-of-two size, then expand if necessary. 768 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 769 EVT NVT = VT.getRoundIntegerType(Context); 770 assert(NVT != VT && "Unable to round integer VT"); 771 LegalizeKind NextStep = getTypeConversion(Context, NVT); 772 // Avoid multi-step promotion. 773 if (NextStep.first == TypePromoteInteger) 774 return NextStep; 775 // Return rounded integer type. 776 return LegalizeKind(TypePromoteInteger, NVT); 777 } 778 779 return LegalizeKind(TypeExpandInteger, 780 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 781 } 782 783 // Handle vector types. 784 unsigned NumElts = VT.getVectorNumElements(); 785 EVT EltVT = VT.getVectorElementType(); 786 787 // Vectors with only one element are always scalarized. 788 if (NumElts == 1) 789 return LegalizeKind(TypeScalarizeVector, EltVT); 790 791 // Try to widen vector elements until the element type is a power of two and 792 // promote it to a legal type later on, for example: 793 // <3 x i8> -> <4 x i8> -> <4 x i32> 794 if (EltVT.isInteger()) { 795 // Vectors with a number of elements that is not a power of two are always 796 // widened, for example <3 x i8> -> <4 x i8>. 797 if (!VT.isPow2VectorType()) { 798 NumElts = (unsigned)NextPowerOf2(NumElts); 799 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 800 return LegalizeKind(TypeWidenVector, NVT); 801 } 802 803 // Examine the element type. 804 LegalizeKind LK = getTypeConversion(Context, EltVT); 805 806 // If type is to be expanded, split the vector. 807 // <4 x i140> -> <2 x i140> 808 if (LK.first == TypeExpandInteger) 809 return LegalizeKind(TypeSplitVector, 810 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 811 812 // Promote the integer element types until a legal vector type is found 813 // or until the element integer type is too big. If a legal type was not 814 // found, fallback to the usual mechanism of widening/splitting the 815 // vector. 816 EVT OldEltVT = EltVT; 817 while (true) { 818 // Increase the bitwidth of the element to the next pow-of-two 819 // (which is greater than 8 bits). 820 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 821 .getRoundIntegerType(Context); 822 823 // Stop trying when getting a non-simple element type. 824 // Note that vector elements may be greater than legal vector element 825 // types. Example: X86 XMM registers hold 64bit element on 32bit 826 // systems. 827 if (!EltVT.isSimple()) 828 break; 829 830 // Build a new vector type and check if it is legal. 831 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 832 // Found a legal promoted vector type. 833 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 834 return LegalizeKind(TypePromoteInteger, 835 EVT::getVectorVT(Context, EltVT, NumElts)); 836 } 837 838 // Reset the type to the unexpanded type if we did not find a legal vector 839 // type with a promoted vector element type. 840 EltVT = OldEltVT; 841 } 842 843 // Try to widen the vector until a legal type is found. 844 // If there is no wider legal type, split the vector. 845 while (true) { 846 // Round up to the next power of 2. 847 NumElts = (unsigned)NextPowerOf2(NumElts); 848 849 // If there is no simple vector type with this many elements then there 850 // cannot be a larger legal vector type. Note that this assumes that 851 // there are no skipped intermediate vector types in the simple types. 852 if (!EltVT.isSimple()) 853 break; 854 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 855 if (LargerVector == MVT()) 856 break; 857 858 // If this type is legal then widen the vector. 859 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 860 return LegalizeKind(TypeWidenVector, LargerVector); 861 } 862 863 // Widen odd vectors to next power of two. 864 if (!VT.isPow2VectorType()) { 865 EVT NVT = VT.getPow2VectorType(Context); 866 return LegalizeKind(TypeWidenVector, NVT); 867 } 868 869 // Vectors with illegal element types are expanded. 870 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 871 return LegalizeKind(TypeSplitVector, NVT); 872 } 873 874 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 875 unsigned &NumIntermediates, 876 MVT &RegisterVT, 877 TargetLoweringBase *TLI) { 878 // Figure out the right, legal destination reg to copy into. 879 unsigned NumElts = VT.getVectorNumElements(); 880 MVT EltTy = VT.getVectorElementType(); 881 882 unsigned NumVectorRegs = 1; 883 884 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 885 // could break down into LHS/RHS like LegalizeDAG does. 886 if (!isPowerOf2_32(NumElts)) { 887 NumVectorRegs = NumElts; 888 NumElts = 1; 889 } 890 891 // Divide the input until we get to a supported size. This will always 892 // end with a scalar if the target doesn't support vectors. 893 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 894 NumElts >>= 1; 895 NumVectorRegs <<= 1; 896 } 897 898 NumIntermediates = NumVectorRegs; 899 900 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 901 if (!TLI->isTypeLegal(NewVT)) 902 NewVT = EltTy; 903 IntermediateVT = NewVT; 904 905 unsigned NewVTSize = NewVT.getSizeInBits(); 906 907 // Convert sizes such as i33 to i64. 908 if (!isPowerOf2_32(NewVTSize)) 909 NewVTSize = NextPowerOf2(NewVTSize); 910 911 MVT DestVT = TLI->getRegisterType(NewVT); 912 RegisterVT = DestVT; 913 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 914 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 915 916 // Otherwise, promotion or legal types use the same number of registers as 917 // the vector decimated to the appropriate level. 918 return NumVectorRegs; 919 } 920 921 /// isLegalRC - Return true if the value types that can be represented by the 922 /// specified register class are all legal. 923 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 924 const TargetRegisterClass &RC) const { 925 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 926 if (isTypeLegal(*I)) 927 return true; 928 return false; 929 } 930 931 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 932 /// sequence of memory operands that is recognized by PrologEpilogInserter. 933 MachineBasicBlock * 934 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 935 MachineBasicBlock *MBB) const { 936 MachineInstr *MI = &InitialMI; 937 MachineFunction &MF = *MI->getMF(); 938 MachineFrameInfo &MFI = MF.getFrameInfo(); 939 940 // We're handling multiple types of operands here: 941 // PATCHPOINT MetaArgs - live-in, read only, direct 942 // STATEPOINT Deopt Spill - live-through, read only, indirect 943 // STATEPOINT Deopt Alloca - live-through, read only, direct 944 // (We're currently conservative and mark the deopt slots read/write in 945 // practice.) 946 // STATEPOINT GC Spill - live-through, read/write, indirect 947 // STATEPOINT GC Alloca - live-through, read/write, direct 948 // The live-in vs live-through is handled already (the live through ones are 949 // all stack slots), but we need to handle the different type of stackmap 950 // operands and memory effects here. 951 952 // MI changes inside this loop as we grow operands. 953 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 954 MachineOperand &MO = MI->getOperand(OperIdx); 955 if (!MO.isFI()) 956 continue; 957 958 // foldMemoryOperand builds a new MI after replacing a single FI operand 959 // with the canonical set of five x86 addressing-mode operands. 960 int FI = MO.getIndex(); 961 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 962 963 // Copy operands before the frame-index. 964 for (unsigned i = 0; i < OperIdx; ++i) 965 MIB.add(MI->getOperand(i)); 966 // Add frame index operands recognized by stackmaps.cpp 967 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 968 // indirect-mem-ref tag, size, #FI, offset. 969 // Used for spills inserted by StatepointLowering. This codepath is not 970 // used for patchpoints/stackmaps at all, for these spilling is done via 971 // foldMemoryOperand callback only. 972 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 973 MIB.addImm(StackMaps::IndirectMemRefOp); 974 MIB.addImm(MFI.getObjectSize(FI)); 975 MIB.add(MI->getOperand(OperIdx)); 976 MIB.addImm(0); 977 } else { 978 // direct-mem-ref tag, #FI, offset. 979 // Used by patchpoint, and direct alloca arguments to statepoints 980 MIB.addImm(StackMaps::DirectMemRefOp); 981 MIB.add(MI->getOperand(OperIdx)); 982 MIB.addImm(0); 983 } 984 // Copy the operands after the frame index. 985 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 986 MIB.add(MI->getOperand(i)); 987 988 // Inherit previous memory operands. 989 MIB.cloneMemRefs(*MI); 990 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 991 992 // Add a new memory operand for this FI. 993 assert(MFI.getObjectOffset(FI) != -1); 994 995 auto Flags = MachineMemOperand::MOLoad; 996 if (MI->getOpcode() == TargetOpcode::STATEPOINT) { 997 Flags |= MachineMemOperand::MOStore; 998 Flags |= MachineMemOperand::MOVolatile; 999 } 1000 MachineMemOperand *MMO = MF.getMachineMemOperand( 1001 MachinePointerInfo::getFixedStack(MF, FI), Flags, 1002 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 1003 MIB->addMemOperand(MF, MMO); 1004 1005 // Replace the instruction and update the operand index. 1006 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1007 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1008 MI->eraseFromParent(); 1009 MI = MIB; 1010 } 1011 return MBB; 1012 } 1013 1014 MachineBasicBlock * 1015 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1016 MachineBasicBlock *MBB) const { 1017 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1018 "Called emitXRayCustomEvent on the wrong MI!"); 1019 auto &MF = *MI.getMF(); 1020 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1021 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1022 MIB.add(MI.getOperand(OpIdx)); 1023 1024 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1025 MI.eraseFromParent(); 1026 return MBB; 1027 } 1028 1029 MachineBasicBlock * 1030 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1031 MachineBasicBlock *MBB) const { 1032 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1033 "Called emitXRayTypedEvent on the wrong MI!"); 1034 auto &MF = *MI.getMF(); 1035 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1036 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1037 MIB.add(MI.getOperand(OpIdx)); 1038 1039 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1040 MI.eraseFromParent(); 1041 return MBB; 1042 } 1043 1044 /// findRepresentativeClass - Return the largest legal super-reg register class 1045 /// of the register class for the specified type and its associated "cost". 1046 // This function is in TargetLowering because it uses RegClassForVT which would 1047 // need to be moved to TargetRegisterInfo and would necessitate moving 1048 // isTypeLegal over as well - a massive change that would just require 1049 // TargetLowering having a TargetRegisterInfo class member that it would use. 1050 std::pair<const TargetRegisterClass *, uint8_t> 1051 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1052 MVT VT) const { 1053 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1054 if (!RC) 1055 return std::make_pair(RC, 0); 1056 1057 // Compute the set of all super-register classes. 1058 BitVector SuperRegRC(TRI->getNumRegClasses()); 1059 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1060 SuperRegRC.setBitsInMask(RCI.getMask()); 1061 1062 // Find the first legal register class with the largest spill size. 1063 const TargetRegisterClass *BestRC = RC; 1064 for (unsigned i : SuperRegRC.set_bits()) { 1065 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1066 // We want the largest possible spill size. 1067 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1068 continue; 1069 if (!isLegalRC(*TRI, *SuperRC)) 1070 continue; 1071 BestRC = SuperRC; 1072 } 1073 return std::make_pair(BestRC, 1); 1074 } 1075 1076 /// computeRegisterProperties - Once all of the register classes are added, 1077 /// this allows us to compute derived properties we expose. 1078 void TargetLoweringBase::computeRegisterProperties( 1079 const TargetRegisterInfo *TRI) { 1080 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1081 "Too many value types for ValueTypeActions to hold!"); 1082 1083 // Everything defaults to needing one register. 1084 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1085 NumRegistersForVT[i] = 1; 1086 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1087 } 1088 // ...except isVoid, which doesn't need any registers. 1089 NumRegistersForVT[MVT::isVoid] = 0; 1090 1091 // Find the largest integer register class. 1092 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1093 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1094 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1095 1096 // Every integer value type larger than this largest register takes twice as 1097 // many registers to represent as the previous ValueType. 1098 for (unsigned ExpandedReg = LargestIntReg + 1; 1099 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1100 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1101 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1102 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1103 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1104 TypeExpandInteger); 1105 } 1106 1107 // Inspect all of the ValueType's smaller than the largest integer 1108 // register to see which ones need promotion. 1109 unsigned LegalIntReg = LargestIntReg; 1110 for (unsigned IntReg = LargestIntReg - 1; 1111 IntReg >= (unsigned)MVT::i1; --IntReg) { 1112 MVT IVT = (MVT::SimpleValueType)IntReg; 1113 if (isTypeLegal(IVT)) { 1114 LegalIntReg = IntReg; 1115 } else { 1116 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1117 (MVT::SimpleValueType)LegalIntReg; 1118 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1119 } 1120 } 1121 1122 // ppcf128 type is really two f64's. 1123 if (!isTypeLegal(MVT::ppcf128)) { 1124 if (isTypeLegal(MVT::f64)) { 1125 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1126 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1127 TransformToType[MVT::ppcf128] = MVT::f64; 1128 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1129 } else { 1130 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1131 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1132 TransformToType[MVT::ppcf128] = MVT::i128; 1133 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1134 } 1135 } 1136 1137 // Decide how to handle f128. If the target does not have native f128 support, 1138 // expand it to i128 and we will be generating soft float library calls. 1139 if (!isTypeLegal(MVT::f128)) { 1140 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1141 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1142 TransformToType[MVT::f128] = MVT::i128; 1143 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1144 } 1145 1146 // Decide how to handle f64. If the target does not have native f64 support, 1147 // expand it to i64 and we will be generating soft float library calls. 1148 if (!isTypeLegal(MVT::f64)) { 1149 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1150 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1151 TransformToType[MVT::f64] = MVT::i64; 1152 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1153 } 1154 1155 // Decide how to handle f32. If the target does not have native f32 support, 1156 // expand it to i32 and we will be generating soft float library calls. 1157 if (!isTypeLegal(MVT::f32)) { 1158 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1159 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1160 TransformToType[MVT::f32] = MVT::i32; 1161 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1162 } 1163 1164 // Decide how to handle f16. If the target does not have native f16 support, 1165 // promote it to f32, because there are no f16 library calls (except for 1166 // conversions). 1167 if (!isTypeLegal(MVT::f16)) { 1168 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1169 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1170 TransformToType[MVT::f16] = MVT::f32; 1171 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1172 } 1173 1174 // Loop over all of the vector value types to see which need transformations. 1175 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1176 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1177 MVT VT = (MVT::SimpleValueType) i; 1178 if (isTypeLegal(VT)) 1179 continue; 1180 1181 MVT EltVT = VT.getVectorElementType(); 1182 unsigned NElts = VT.getVectorNumElements(); 1183 bool IsLegalWiderType = false; 1184 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1185 switch (PreferredAction) { 1186 case TypePromoteInteger: 1187 // Try to promote the elements of integer vectors. If no legal 1188 // promotion was found, fall through to the widen-vector method. 1189 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1190 MVT SVT = (MVT::SimpleValueType) nVT; 1191 // Promote vectors of integers to vectors with the same number 1192 // of elements, with a wider element type. 1193 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1194 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1195 TransformToType[i] = SVT; 1196 RegisterTypeForVT[i] = SVT; 1197 NumRegistersForVT[i] = 1; 1198 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1199 IsLegalWiderType = true; 1200 break; 1201 } 1202 } 1203 if (IsLegalWiderType) 1204 break; 1205 LLVM_FALLTHROUGH; 1206 1207 case TypeWidenVector: 1208 // Try to widen the vector. 1209 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1210 MVT SVT = (MVT::SimpleValueType) nVT; 1211 if (SVT.getVectorElementType() == EltVT 1212 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1213 TransformToType[i] = SVT; 1214 RegisterTypeForVT[i] = SVT; 1215 NumRegistersForVT[i] = 1; 1216 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1217 IsLegalWiderType = true; 1218 break; 1219 } 1220 } 1221 if (IsLegalWiderType) 1222 break; 1223 LLVM_FALLTHROUGH; 1224 1225 case TypeSplitVector: 1226 case TypeScalarizeVector: { 1227 MVT IntermediateVT; 1228 MVT RegisterVT; 1229 unsigned NumIntermediates; 1230 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1231 NumIntermediates, RegisterVT, this); 1232 RegisterTypeForVT[i] = RegisterVT; 1233 1234 MVT NVT = VT.getPow2VectorType(); 1235 if (NVT == VT) { 1236 // Type is already a power of 2. The default action is to split. 1237 TransformToType[i] = MVT::Other; 1238 if (PreferredAction == TypeScalarizeVector) 1239 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1240 else if (PreferredAction == TypeSplitVector) 1241 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1242 else 1243 // Set type action according to the number of elements. 1244 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1245 : TypeSplitVector); 1246 } else { 1247 TransformToType[i] = NVT; 1248 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1249 } 1250 break; 1251 } 1252 default: 1253 llvm_unreachable("Unknown vector legalization action!"); 1254 } 1255 } 1256 1257 // Determine the 'representative' register class for each value type. 1258 // An representative register class is the largest (meaning one which is 1259 // not a sub-register class / subreg register class) legal register class for 1260 // a group of value types. For example, on i386, i8, i16, and i32 1261 // representative would be GR32; while on x86_64 it's GR64. 1262 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1263 const TargetRegisterClass* RRC; 1264 uint8_t Cost; 1265 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1266 RepRegClassForVT[i] = RRC; 1267 RepRegClassCostForVT[i] = Cost; 1268 } 1269 } 1270 1271 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1272 EVT VT) const { 1273 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1274 return getPointerTy(DL).SimpleTy; 1275 } 1276 1277 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1278 return MVT::i32; // return the default value 1279 } 1280 1281 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1282 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1283 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1284 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1285 /// 1286 /// This method returns the number of registers needed, and the VT for each 1287 /// register. It also returns the VT and quantity of the intermediate values 1288 /// before they are promoted/expanded. 1289 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1290 EVT &IntermediateVT, 1291 unsigned &NumIntermediates, 1292 MVT &RegisterVT) const { 1293 unsigned NumElts = VT.getVectorNumElements(); 1294 1295 // If there is a wider vector type with the same element type as this one, 1296 // or a promoted vector type that has the same number of elements which 1297 // are wider, then we should convert to that legal vector type. 1298 // This handles things like <2 x float> -> <4 x float> and 1299 // <4 x i1> -> <4 x i32>. 1300 LegalizeTypeAction TA = getTypeAction(Context, VT); 1301 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1302 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1303 if (isTypeLegal(RegisterEVT)) { 1304 IntermediateVT = RegisterEVT; 1305 RegisterVT = RegisterEVT.getSimpleVT(); 1306 NumIntermediates = 1; 1307 return 1; 1308 } 1309 } 1310 1311 // Figure out the right, legal destination reg to copy into. 1312 EVT EltTy = VT.getVectorElementType(); 1313 1314 unsigned NumVectorRegs = 1; 1315 1316 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1317 // could break down into LHS/RHS like LegalizeDAG does. 1318 if (!isPowerOf2_32(NumElts)) { 1319 NumVectorRegs = NumElts; 1320 NumElts = 1; 1321 } 1322 1323 // Divide the input until we get to a supported size. This will always 1324 // end with a scalar if the target doesn't support vectors. 1325 while (NumElts > 1 && !isTypeLegal( 1326 EVT::getVectorVT(Context, EltTy, NumElts))) { 1327 NumElts >>= 1; 1328 NumVectorRegs <<= 1; 1329 } 1330 1331 NumIntermediates = NumVectorRegs; 1332 1333 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1334 if (!isTypeLegal(NewVT)) 1335 NewVT = EltTy; 1336 IntermediateVT = NewVT; 1337 1338 MVT DestVT = getRegisterType(Context, NewVT); 1339 RegisterVT = DestVT; 1340 unsigned NewVTSize = NewVT.getSizeInBits(); 1341 1342 // Convert sizes such as i33 to i64. 1343 if (!isPowerOf2_32(NewVTSize)) 1344 NewVTSize = NextPowerOf2(NewVTSize); 1345 1346 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1347 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1348 1349 // Otherwise, promotion or legal types use the same number of registers as 1350 // the vector decimated to the appropriate level. 1351 return NumVectorRegs; 1352 } 1353 1354 /// Get the EVTs and ArgFlags collections that represent the legalized return 1355 /// type of the given function. This does not require a DAG or a return value, 1356 /// and is suitable for use before any DAGs for the function are constructed. 1357 /// TODO: Move this out of TargetLowering.cpp. 1358 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1359 AttributeList attr, 1360 SmallVectorImpl<ISD::OutputArg> &Outs, 1361 const TargetLowering &TLI, const DataLayout &DL) { 1362 SmallVector<EVT, 4> ValueVTs; 1363 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1364 unsigned NumValues = ValueVTs.size(); 1365 if (NumValues == 0) return; 1366 1367 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1368 EVT VT = ValueVTs[j]; 1369 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1370 1371 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1372 ExtendKind = ISD::SIGN_EXTEND; 1373 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1374 ExtendKind = ISD::ZERO_EXTEND; 1375 1376 // FIXME: C calling convention requires the return type to be promoted to 1377 // at least 32-bit. But this is not necessary for non-C calling 1378 // conventions. The frontend should mark functions whose return values 1379 // require promoting with signext or zeroext attributes. 1380 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1381 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1382 if (VT.bitsLT(MinVT)) 1383 VT = MinVT; 1384 } 1385 1386 unsigned NumParts = 1387 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1388 MVT PartVT = 1389 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1390 1391 // 'inreg' on function refers to return value 1392 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1393 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1394 Flags.setInReg(); 1395 1396 // Propagate extension type if any 1397 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1398 Flags.setSExt(); 1399 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1400 Flags.setZExt(); 1401 1402 for (unsigned i = 0; i < NumParts; ++i) 1403 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0)); 1404 } 1405 } 1406 1407 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1408 /// function arguments in the caller parameter area. This is the actual 1409 /// alignment, not its logarithm. 1410 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1411 const DataLayout &DL) const { 1412 return DL.getABITypeAlignment(Ty); 1413 } 1414 1415 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1416 const DataLayout &DL, EVT VT, 1417 unsigned AddrSpace, 1418 unsigned Alignment, 1419 bool *Fast) const { 1420 // Check if the specified alignment is sufficient based on the data layout. 1421 // TODO: While using the data layout works in practice, a better solution 1422 // would be to implement this check directly (make this a virtual function). 1423 // For example, the ABI alignment may change based on software platform while 1424 // this function should only be affected by hardware implementation. 1425 Type *Ty = VT.getTypeForEVT(Context); 1426 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1427 // Assume that an access that meets the ABI-specified alignment is fast. 1428 if (Fast != nullptr) 1429 *Fast = true; 1430 return true; 1431 } 1432 1433 // This is a misaligned access. 1434 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast); 1435 } 1436 1437 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1438 return BranchProbability(MinPercentageForPredictableBranch, 100); 1439 } 1440 1441 //===----------------------------------------------------------------------===// 1442 // TargetTransformInfo Helpers 1443 //===----------------------------------------------------------------------===// 1444 1445 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1446 enum InstructionOpcodes { 1447 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1448 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1449 #include "llvm/IR/Instruction.def" 1450 }; 1451 switch (static_cast<InstructionOpcodes>(Opcode)) { 1452 case Ret: return 0; 1453 case Br: return 0; 1454 case Switch: return 0; 1455 case IndirectBr: return 0; 1456 case Invoke: return 0; 1457 case Resume: return 0; 1458 case Unreachable: return 0; 1459 case CleanupRet: return 0; 1460 case CatchRet: return 0; 1461 case CatchPad: return 0; 1462 case CatchSwitch: return 0; 1463 case CleanupPad: return 0; 1464 case FNeg: return ISD::FNEG; 1465 case Add: return ISD::ADD; 1466 case FAdd: return ISD::FADD; 1467 case Sub: return ISD::SUB; 1468 case FSub: return ISD::FSUB; 1469 case Mul: return ISD::MUL; 1470 case FMul: return ISD::FMUL; 1471 case UDiv: return ISD::UDIV; 1472 case SDiv: return ISD::SDIV; 1473 case FDiv: return ISD::FDIV; 1474 case URem: return ISD::UREM; 1475 case SRem: return ISD::SREM; 1476 case FRem: return ISD::FREM; 1477 case Shl: return ISD::SHL; 1478 case LShr: return ISD::SRL; 1479 case AShr: return ISD::SRA; 1480 case And: return ISD::AND; 1481 case Or: return ISD::OR; 1482 case Xor: return ISD::XOR; 1483 case Alloca: return 0; 1484 case Load: return ISD::LOAD; 1485 case Store: return ISD::STORE; 1486 case GetElementPtr: return 0; 1487 case Fence: return 0; 1488 case AtomicCmpXchg: return 0; 1489 case AtomicRMW: return 0; 1490 case Trunc: return ISD::TRUNCATE; 1491 case ZExt: return ISD::ZERO_EXTEND; 1492 case SExt: return ISD::SIGN_EXTEND; 1493 case FPToUI: return ISD::FP_TO_UINT; 1494 case FPToSI: return ISD::FP_TO_SINT; 1495 case UIToFP: return ISD::UINT_TO_FP; 1496 case SIToFP: return ISD::SINT_TO_FP; 1497 case FPTrunc: return ISD::FP_ROUND; 1498 case FPExt: return ISD::FP_EXTEND; 1499 case PtrToInt: return ISD::BITCAST; 1500 case IntToPtr: return ISD::BITCAST; 1501 case BitCast: return ISD::BITCAST; 1502 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1503 case ICmp: return ISD::SETCC; 1504 case FCmp: return ISD::SETCC; 1505 case PHI: return 0; 1506 case Call: return 0; 1507 case Select: return ISD::SELECT; 1508 case UserOp1: return 0; 1509 case UserOp2: return 0; 1510 case VAArg: return 0; 1511 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1512 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1513 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1514 case ExtractValue: return ISD::MERGE_VALUES; 1515 case InsertValue: return ISD::MERGE_VALUES; 1516 case LandingPad: return 0; 1517 } 1518 1519 llvm_unreachable("Unknown instruction type encountered!"); 1520 } 1521 1522 std::pair<int, MVT> 1523 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1524 Type *Ty) const { 1525 LLVMContext &C = Ty->getContext(); 1526 EVT MTy = getValueType(DL, Ty); 1527 1528 int Cost = 1; 1529 // We keep legalizing the type until we find a legal kind. We assume that 1530 // the only operation that costs anything is the split. After splitting 1531 // we need to handle two types. 1532 while (true) { 1533 LegalizeKind LK = getTypeConversion(C, MTy); 1534 1535 if (LK.first == TypeLegal) 1536 return std::make_pair(Cost, MTy.getSimpleVT()); 1537 1538 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1539 Cost *= 2; 1540 1541 // Do not loop with f128 type. 1542 if (MTy == LK.second) 1543 return std::make_pair(Cost, MTy.getSimpleVT()); 1544 1545 // Keep legalizing the type. 1546 MTy = LK.second; 1547 } 1548 } 1549 1550 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1551 bool UseTLS) const { 1552 // compiler-rt provides a variable with a magic name. Targets that do not 1553 // link with compiler-rt may also provide such a variable. 1554 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1555 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1556 auto UnsafeStackPtr = 1557 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1558 1559 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1560 1561 if (!UnsafeStackPtr) { 1562 auto TLSModel = UseTLS ? 1563 GlobalValue::InitialExecTLSModel : 1564 GlobalValue::NotThreadLocal; 1565 // The global variable is not defined yet, define it ourselves. 1566 // We use the initial-exec TLS model because we do not support the 1567 // variable living anywhere other than in the main executable. 1568 UnsafeStackPtr = new GlobalVariable( 1569 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1570 UnsafeStackPtrVar, nullptr, TLSModel); 1571 } else { 1572 // The variable exists, check its type and attributes. 1573 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1574 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1575 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1576 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1577 (UseTLS ? "" : "not ") + "be thread-local"); 1578 } 1579 return UnsafeStackPtr; 1580 } 1581 1582 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1583 if (!TM.getTargetTriple().isAndroid()) 1584 return getDefaultSafeStackPointerLocation(IRB, true); 1585 1586 // Android provides a libc function to retrieve the address of the current 1587 // thread's unsafe stack pointer. 1588 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1589 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1590 Value *Fn = M->getOrInsertFunction("__safestack_pointer_address", 1591 StackPtrTy->getPointerTo(0)); 1592 return IRB.CreateCall(Fn); 1593 } 1594 1595 //===----------------------------------------------------------------------===// 1596 // Loop Strength Reduction hooks 1597 //===----------------------------------------------------------------------===// 1598 1599 /// isLegalAddressingMode - Return true if the addressing mode represented 1600 /// by AM is legal for this target, for a load/store of the specified type. 1601 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1602 const AddrMode &AM, Type *Ty, 1603 unsigned AS, Instruction *I) const { 1604 // The default implementation of this implements a conservative RISCy, r+r and 1605 // r+i addr mode. 1606 1607 // Allows a sign-extended 16-bit immediate field. 1608 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1609 return false; 1610 1611 // No global is ever allowed as a base. 1612 if (AM.BaseGV) 1613 return false; 1614 1615 // Only support r+r, 1616 switch (AM.Scale) { 1617 case 0: // "r+i" or just "i", depending on HasBaseReg. 1618 break; 1619 case 1: 1620 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1621 return false; 1622 // Otherwise we have r+r or r+i. 1623 break; 1624 case 2: 1625 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1626 return false; 1627 // Allow 2*r as r+r. 1628 break; 1629 default: // Don't allow n * r 1630 return false; 1631 } 1632 1633 return true; 1634 } 1635 1636 //===----------------------------------------------------------------------===// 1637 // Stack Protector 1638 //===----------------------------------------------------------------------===// 1639 1640 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1641 // so that SelectionDAG handle SSP. 1642 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1643 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1644 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1645 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1646 return M.getOrInsertGlobal("__guard_local", PtrTy); 1647 } 1648 return nullptr; 1649 } 1650 1651 // Currently only support "standard" __stack_chk_guard. 1652 // TODO: add LOAD_STACK_GUARD support. 1653 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1654 if (!M.getNamedValue("__stack_chk_guard")) 1655 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1656 GlobalVariable::ExternalLinkage, 1657 nullptr, "__stack_chk_guard"); 1658 } 1659 1660 // Currently only support "standard" __stack_chk_guard. 1661 // TODO: add LOAD_STACK_GUARD support. 1662 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1663 return M.getNamedValue("__stack_chk_guard"); 1664 } 1665 1666 Value *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1667 return nullptr; 1668 } 1669 1670 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1671 return MinimumJumpTableEntries; 1672 } 1673 1674 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1675 MinimumJumpTableEntries = Val; 1676 } 1677 1678 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1679 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1680 } 1681 1682 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1683 return MaximumJumpTableSize; 1684 } 1685 1686 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1687 MaximumJumpTableSize = Val; 1688 } 1689 1690 //===----------------------------------------------------------------------===// 1691 // Reciprocal Estimates 1692 //===----------------------------------------------------------------------===// 1693 1694 /// Get the reciprocal estimate attribute string for a function that will 1695 /// override the target defaults. 1696 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1697 const Function &F = MF.getFunction(); 1698 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1699 } 1700 1701 /// Construct a string for the given reciprocal operation of the given type. 1702 /// This string should match the corresponding option to the front-end's 1703 /// "-mrecip" flag assuming those strings have been passed through in an 1704 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1705 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1706 std::string Name = VT.isVector() ? "vec-" : ""; 1707 1708 Name += IsSqrt ? "sqrt" : "div"; 1709 1710 // TODO: Handle "half" or other float types? 1711 if (VT.getScalarType() == MVT::f64) { 1712 Name += "d"; 1713 } else { 1714 assert(VT.getScalarType() == MVT::f32 && 1715 "Unexpected FP type for reciprocal estimate"); 1716 Name += "f"; 1717 } 1718 1719 return Name; 1720 } 1721 1722 /// Return the character position and value (a single numeric character) of a 1723 /// customized refinement operation in the input string if it exists. Return 1724 /// false if there is no customized refinement step count. 1725 static bool parseRefinementStep(StringRef In, size_t &Position, 1726 uint8_t &Value) { 1727 const char RefStepToken = ':'; 1728 Position = In.find(RefStepToken); 1729 if (Position == StringRef::npos) 1730 return false; 1731 1732 StringRef RefStepString = In.substr(Position + 1); 1733 // Allow exactly one numeric character for the additional refinement 1734 // step parameter. 1735 if (RefStepString.size() == 1) { 1736 char RefStepChar = RefStepString[0]; 1737 if (RefStepChar >= '0' && RefStepChar <= '9') { 1738 Value = RefStepChar - '0'; 1739 return true; 1740 } 1741 } 1742 report_fatal_error("Invalid refinement step for -recip."); 1743 } 1744 1745 /// For the input attribute string, return one of the ReciprocalEstimate enum 1746 /// status values (enabled, disabled, or not specified) for this operation on 1747 /// the specified data type. 1748 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1749 if (Override.empty()) 1750 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1751 1752 SmallVector<StringRef, 4> OverrideVector; 1753 Override.split(OverrideVector, ','); 1754 unsigned NumArgs = OverrideVector.size(); 1755 1756 // Check if "all", "none", or "default" was specified. 1757 if (NumArgs == 1) { 1758 // Look for an optional setting of the number of refinement steps needed 1759 // for this type of reciprocal operation. 1760 size_t RefPos; 1761 uint8_t RefSteps; 1762 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1763 // Split the string for further processing. 1764 Override = Override.substr(0, RefPos); 1765 } 1766 1767 // All reciprocal types are enabled. 1768 if (Override == "all") 1769 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1770 1771 // All reciprocal types are disabled. 1772 if (Override == "none") 1773 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1774 1775 // Target defaults for enablement are used. 1776 if (Override == "default") 1777 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1778 } 1779 1780 // The attribute string may omit the size suffix ('f'/'d'). 1781 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1782 std::string VTNameNoSize = VTName; 1783 VTNameNoSize.pop_back(); 1784 static const char DisabledPrefix = '!'; 1785 1786 for (StringRef RecipType : OverrideVector) { 1787 size_t RefPos; 1788 uint8_t RefSteps; 1789 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1790 RecipType = RecipType.substr(0, RefPos); 1791 1792 // Ignore the disablement token for string matching. 1793 bool IsDisabled = RecipType[0] == DisabledPrefix; 1794 if (IsDisabled) 1795 RecipType = RecipType.substr(1); 1796 1797 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1798 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1799 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1800 } 1801 1802 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1803 } 1804 1805 /// For the input attribute string, return the customized refinement step count 1806 /// for this operation on the specified data type. If the step count does not 1807 /// exist, return the ReciprocalEstimate enum value for unspecified. 1808 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1809 if (Override.empty()) 1810 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1811 1812 SmallVector<StringRef, 4> OverrideVector; 1813 Override.split(OverrideVector, ','); 1814 unsigned NumArgs = OverrideVector.size(); 1815 1816 // Check if "all", "default", or "none" was specified. 1817 if (NumArgs == 1) { 1818 // Look for an optional setting of the number of refinement steps needed 1819 // for this type of reciprocal operation. 1820 size_t RefPos; 1821 uint8_t RefSteps; 1822 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1823 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1824 1825 // Split the string for further processing. 1826 Override = Override.substr(0, RefPos); 1827 assert(Override != "none" && 1828 "Disabled reciprocals, but specifed refinement steps?"); 1829 1830 // If this is a general override, return the specified number of steps. 1831 if (Override == "all" || Override == "default") 1832 return RefSteps; 1833 } 1834 1835 // The attribute string may omit the size suffix ('f'/'d'). 1836 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1837 std::string VTNameNoSize = VTName; 1838 VTNameNoSize.pop_back(); 1839 1840 for (StringRef RecipType : OverrideVector) { 1841 size_t RefPos; 1842 uint8_t RefSteps; 1843 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1844 continue; 1845 1846 RecipType = RecipType.substr(0, RefPos); 1847 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1848 return RefSteps; 1849 } 1850 1851 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1852 } 1853 1854 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1855 MachineFunction &MF) const { 1856 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1857 } 1858 1859 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1860 MachineFunction &MF) const { 1861 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1862 } 1863 1864 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1865 MachineFunction &MF) const { 1866 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1867 } 1868 1869 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1870 MachineFunction &MF) const { 1871 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1872 } 1873 1874 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1875 MF.getRegInfo().freezeReservedRegs(MF); 1876 } 1877