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(UINT_MAX), cl::Hidden, 77 cl::desc("Set maximum size of jump tables.")); 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 StackPointerRegisterToSaveRestore = 0; 549 BooleanContents = UndefinedBooleanContent; 550 BooleanFloatContents = UndefinedBooleanContent; 551 BooleanVectorContents = UndefinedBooleanContent; 552 SchedPreferenceInfo = Sched::ILP; 553 JumpBufSize = 0; 554 JumpBufAlignment = 0; 555 MinFunctionAlignment = 0; 556 PrefFunctionAlignment = 0; 557 PrefLoopAlignment = 0; 558 GatherAllAliasesMaxDepth = 18; 559 MinStackArgumentAlignment = 1; 560 // TODO: the default will be switched to 0 in the next commit, along 561 // with the Target-specific changes necessary. 562 MaxAtomicSizeInBitsSupported = 1024; 563 564 MinCmpXchgSizeInBits = 0; 565 SupportsUnalignedAtomics = false; 566 567 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 568 569 InitLibcalls(TM.getTargetTriple()); 570 InitCmpLibcallCCs(CmpLibcallCCs); 571 } 572 573 void TargetLoweringBase::initActions() { 574 // All operations default to being supported. 575 memset(OpActions, 0, sizeof(OpActions)); 576 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 577 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 578 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 579 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 580 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 581 std::fill(std::begin(TargetDAGCombineArray), 582 std::end(TargetDAGCombineArray), 0); 583 584 for (MVT VT : MVT::fp_valuetypes()) { 585 MVT IntVT = MVT::getIntegerVT(VT.getSizeInBits()); 586 if (IntVT.isValid()) { 587 setOperationAction(ISD::ATOMIC_SWAP, VT, Promote); 588 AddPromotedToType(ISD::ATOMIC_SWAP, VT, IntVT); 589 } 590 } 591 592 // Set default actions for various operations. 593 for (MVT VT : MVT::all_valuetypes()) { 594 // Default all indexed load / store to expand. 595 for (unsigned IM = (unsigned)ISD::PRE_INC; 596 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 597 setIndexedLoadAction(IM, VT, Expand); 598 setIndexedStoreAction(IM, VT, Expand); 599 } 600 601 // Most backends expect to see the node which just returns the value loaded. 602 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 603 604 // These operations default to expand. 605 setOperationAction(ISD::FGETSIGN, VT, Expand); 606 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 607 setOperationAction(ISD::FMINNUM, VT, Expand); 608 setOperationAction(ISD::FMAXNUM, VT, Expand); 609 setOperationAction(ISD::FMINNUM_IEEE, VT, Expand); 610 setOperationAction(ISD::FMAXNUM_IEEE, VT, Expand); 611 setOperationAction(ISD::FMINIMUM, VT, Expand); 612 setOperationAction(ISD::FMAXIMUM, VT, Expand); 613 setOperationAction(ISD::FMAD, VT, Expand); 614 setOperationAction(ISD::SMIN, VT, Expand); 615 setOperationAction(ISD::SMAX, VT, Expand); 616 setOperationAction(ISD::UMIN, VT, Expand); 617 setOperationAction(ISD::UMAX, VT, Expand); 618 setOperationAction(ISD::ABS, VT, Expand); 619 setOperationAction(ISD::FSHL, VT, Expand); 620 setOperationAction(ISD::FSHR, VT, Expand); 621 setOperationAction(ISD::SADDSAT, VT, Expand); 622 setOperationAction(ISD::UADDSAT, VT, Expand); 623 setOperationAction(ISD::SSUBSAT, VT, Expand); 624 setOperationAction(ISD::USUBSAT, VT, Expand); 625 setOperationAction(ISD::SMULFIX, VT, Expand); 626 setOperationAction(ISD::SMULFIXSAT, 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 // Vector reduction default to expand. 670 setOperationAction(ISD::VECREDUCE_FADD, VT, Expand); 671 setOperationAction(ISD::VECREDUCE_FMUL, VT, Expand); 672 setOperationAction(ISD::VECREDUCE_ADD, VT, Expand); 673 setOperationAction(ISD::VECREDUCE_MUL, VT, Expand); 674 setOperationAction(ISD::VECREDUCE_AND, VT, Expand); 675 setOperationAction(ISD::VECREDUCE_OR, VT, Expand); 676 setOperationAction(ISD::VECREDUCE_XOR, VT, Expand); 677 setOperationAction(ISD::VECREDUCE_SMAX, VT, Expand); 678 setOperationAction(ISD::VECREDUCE_SMIN, VT, Expand); 679 setOperationAction(ISD::VECREDUCE_UMAX, VT, Expand); 680 setOperationAction(ISD::VECREDUCE_UMIN, VT, Expand); 681 setOperationAction(ISD::VECREDUCE_FMAX, VT, Expand); 682 setOperationAction(ISD::VECREDUCE_FMIN, VT, Expand); 683 } 684 685 // Most targets ignore the @llvm.prefetch intrinsic. 686 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 687 688 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 689 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 690 691 // ConstantFP nodes default to expand. Targets can either change this to 692 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 693 // to optimize expansions for certain constants. 694 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 695 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 696 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 697 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 698 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 699 700 // These library functions default to expand. 701 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 702 setOperationAction(ISD::FCBRT, VT, Expand); 703 setOperationAction(ISD::FLOG , VT, Expand); 704 setOperationAction(ISD::FLOG2, VT, Expand); 705 setOperationAction(ISD::FLOG10, VT, Expand); 706 setOperationAction(ISD::FEXP , VT, Expand); 707 setOperationAction(ISD::FEXP2, VT, Expand); 708 setOperationAction(ISD::FFLOOR, VT, Expand); 709 setOperationAction(ISD::FNEARBYINT, VT, Expand); 710 setOperationAction(ISD::FCEIL, VT, Expand); 711 setOperationAction(ISD::FRINT, VT, Expand); 712 setOperationAction(ISD::FTRUNC, VT, Expand); 713 setOperationAction(ISD::FROUND, VT, Expand); 714 setOperationAction(ISD::LROUND, VT, Expand); 715 setOperationAction(ISD::LLROUND, VT, Expand); 716 } 717 718 // Default ISD::TRAP to expand (which turns it into abort). 719 setOperationAction(ISD::TRAP, MVT::Other, Expand); 720 721 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 722 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 723 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 724 } 725 726 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 727 EVT) const { 728 return MVT::getIntegerVT(DL.getPointerSizeInBits(0)); 729 } 730 731 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, const DataLayout &DL, 732 bool LegalTypes) const { 733 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 734 if (LHSTy.isVector()) 735 return LHSTy; 736 return LegalTypes ? getScalarShiftAmountTy(DL, LHSTy) 737 : getPointerTy(DL); 738 } 739 740 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 741 assert(isTypeLegal(VT)); 742 switch (Op) { 743 default: 744 return false; 745 case ISD::SDIV: 746 case ISD::UDIV: 747 case ISD::SREM: 748 case ISD::UREM: 749 return true; 750 } 751 } 752 753 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 754 // If the command-line option was specified, ignore this request. 755 if (!JumpIsExpensiveOverride.getNumOccurrences()) 756 JumpIsExpensive = isExpensive; 757 } 758 759 TargetLoweringBase::LegalizeKind 760 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 761 // If this is a simple type, use the ComputeRegisterProp mechanism. 762 if (VT.isSimple()) { 763 MVT SVT = VT.getSimpleVT(); 764 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 765 MVT NVT = TransformToType[SVT.SimpleTy]; 766 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 767 768 assert((LA == TypeLegal || LA == TypeSoftenFloat || 769 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) && 770 "Promote may not follow Expand or Promote"); 771 772 if (LA == TypeSplitVector) 773 return LegalizeKind(LA, 774 EVT::getVectorVT(Context, SVT.getVectorElementType(), 775 SVT.getVectorNumElements() / 2)); 776 if (LA == TypeScalarizeVector) 777 return LegalizeKind(LA, SVT.getVectorElementType()); 778 return LegalizeKind(LA, NVT); 779 } 780 781 // Handle Extended Scalar Types. 782 if (!VT.isVector()) { 783 assert(VT.isInteger() && "Float types must be simple"); 784 unsigned BitSize = VT.getSizeInBits(); 785 // First promote to a power-of-two size, then expand if necessary. 786 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 787 EVT NVT = VT.getRoundIntegerType(Context); 788 assert(NVT != VT && "Unable to round integer VT"); 789 LegalizeKind NextStep = getTypeConversion(Context, NVT); 790 // Avoid multi-step promotion. 791 if (NextStep.first == TypePromoteInteger) 792 return NextStep; 793 // Return rounded integer type. 794 return LegalizeKind(TypePromoteInteger, NVT); 795 } 796 797 return LegalizeKind(TypeExpandInteger, 798 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 799 } 800 801 // Handle vector types. 802 unsigned NumElts = VT.getVectorNumElements(); 803 EVT EltVT = VT.getVectorElementType(); 804 805 // Vectors with only one element are always scalarized. 806 if (NumElts == 1) 807 return LegalizeKind(TypeScalarizeVector, EltVT); 808 809 // Try to widen vector elements until the element type is a power of two and 810 // promote it to a legal type later on, for example: 811 // <3 x i8> -> <4 x i8> -> <4 x i32> 812 if (EltVT.isInteger()) { 813 // Vectors with a number of elements that is not a power of two are always 814 // widened, for example <3 x i8> -> <4 x i8>. 815 if (!VT.isPow2VectorType()) { 816 NumElts = (unsigned)NextPowerOf2(NumElts); 817 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 818 return LegalizeKind(TypeWidenVector, NVT); 819 } 820 821 // Examine the element type. 822 LegalizeKind LK = getTypeConversion(Context, EltVT); 823 824 // If type is to be expanded, split the vector. 825 // <4 x i140> -> <2 x i140> 826 if (LK.first == TypeExpandInteger) 827 return LegalizeKind(TypeSplitVector, 828 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 829 830 // Promote the integer element types until a legal vector type is found 831 // or until the element integer type is too big. If a legal type was not 832 // found, fallback to the usual mechanism of widening/splitting the 833 // vector. 834 EVT OldEltVT = EltVT; 835 while (true) { 836 // Increase the bitwidth of the element to the next pow-of-two 837 // (which is greater than 8 bits). 838 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 839 .getRoundIntegerType(Context); 840 841 // Stop trying when getting a non-simple element type. 842 // Note that vector elements may be greater than legal vector element 843 // types. Example: X86 XMM registers hold 64bit element on 32bit 844 // systems. 845 if (!EltVT.isSimple()) 846 break; 847 848 // Build a new vector type and check if it is legal. 849 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 850 // Found a legal promoted vector type. 851 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 852 return LegalizeKind(TypePromoteInteger, 853 EVT::getVectorVT(Context, EltVT, NumElts)); 854 } 855 856 // Reset the type to the unexpanded type if we did not find a legal vector 857 // type with a promoted vector element type. 858 EltVT = OldEltVT; 859 } 860 861 // Try to widen the vector until a legal type is found. 862 // If there is no wider legal type, split the vector. 863 while (true) { 864 // Round up to the next power of 2. 865 NumElts = (unsigned)NextPowerOf2(NumElts); 866 867 // If there is no simple vector type with this many elements then there 868 // cannot be a larger legal vector type. Note that this assumes that 869 // there are no skipped intermediate vector types in the simple types. 870 if (!EltVT.isSimple()) 871 break; 872 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 873 if (LargerVector == MVT()) 874 break; 875 876 // If this type is legal then widen the vector. 877 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 878 return LegalizeKind(TypeWidenVector, LargerVector); 879 } 880 881 // Widen odd vectors to next power of two. 882 if (!VT.isPow2VectorType()) { 883 EVT NVT = VT.getPow2VectorType(Context); 884 return LegalizeKind(TypeWidenVector, NVT); 885 } 886 887 // Vectors with illegal element types are expanded. 888 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 889 return LegalizeKind(TypeSplitVector, NVT); 890 } 891 892 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 893 unsigned &NumIntermediates, 894 MVT &RegisterVT, 895 TargetLoweringBase *TLI) { 896 // Figure out the right, legal destination reg to copy into. 897 unsigned NumElts = VT.getVectorNumElements(); 898 MVT EltTy = VT.getVectorElementType(); 899 900 unsigned NumVectorRegs = 1; 901 902 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 903 // could break down into LHS/RHS like LegalizeDAG does. 904 if (!isPowerOf2_32(NumElts)) { 905 NumVectorRegs = NumElts; 906 NumElts = 1; 907 } 908 909 // Divide the input until we get to a supported size. This will always 910 // end with a scalar if the target doesn't support vectors. 911 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 912 NumElts >>= 1; 913 NumVectorRegs <<= 1; 914 } 915 916 NumIntermediates = NumVectorRegs; 917 918 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 919 if (!TLI->isTypeLegal(NewVT)) 920 NewVT = EltTy; 921 IntermediateVT = NewVT; 922 923 unsigned NewVTSize = NewVT.getSizeInBits(); 924 925 // Convert sizes such as i33 to i64. 926 if (!isPowerOf2_32(NewVTSize)) 927 NewVTSize = NextPowerOf2(NewVTSize); 928 929 MVT DestVT = TLI->getRegisterType(NewVT); 930 RegisterVT = DestVT; 931 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 932 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 933 934 // Otherwise, promotion or legal types use the same number of registers as 935 // the vector decimated to the appropriate level. 936 return NumVectorRegs; 937 } 938 939 /// isLegalRC - Return true if the value types that can be represented by the 940 /// specified register class are all legal. 941 bool TargetLoweringBase::isLegalRC(const TargetRegisterInfo &TRI, 942 const TargetRegisterClass &RC) const { 943 for (auto I = TRI.legalclasstypes_begin(RC); *I != MVT::Other; ++I) 944 if (isTypeLegal(*I)) 945 return true; 946 return false; 947 } 948 949 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 950 /// sequence of memory operands that is recognized by PrologEpilogInserter. 951 MachineBasicBlock * 952 TargetLoweringBase::emitPatchPoint(MachineInstr &InitialMI, 953 MachineBasicBlock *MBB) const { 954 MachineInstr *MI = &InitialMI; 955 MachineFunction &MF = *MI->getMF(); 956 MachineFrameInfo &MFI = MF.getFrameInfo(); 957 958 // We're handling multiple types of operands here: 959 // PATCHPOINT MetaArgs - live-in, read only, direct 960 // STATEPOINT Deopt Spill - live-through, read only, indirect 961 // STATEPOINT Deopt Alloca - live-through, read only, direct 962 // (We're currently conservative and mark the deopt slots read/write in 963 // practice.) 964 // STATEPOINT GC Spill - live-through, read/write, indirect 965 // STATEPOINT GC Alloca - live-through, read/write, direct 966 // The live-in vs live-through is handled already (the live through ones are 967 // all stack slots), but we need to handle the different type of stackmap 968 // operands and memory effects here. 969 970 // MI changes inside this loop as we grow operands. 971 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 972 MachineOperand &MO = MI->getOperand(OperIdx); 973 if (!MO.isFI()) 974 continue; 975 976 // foldMemoryOperand builds a new MI after replacing a single FI operand 977 // with the canonical set of five x86 addressing-mode operands. 978 int FI = MO.getIndex(); 979 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 980 981 // Copy operands before the frame-index. 982 for (unsigned i = 0; i < OperIdx; ++i) 983 MIB.add(MI->getOperand(i)); 984 // Add frame index operands recognized by stackmaps.cpp 985 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 986 // indirect-mem-ref tag, size, #FI, offset. 987 // Used for spills inserted by StatepointLowering. This codepath is not 988 // used for patchpoints/stackmaps at all, for these spilling is done via 989 // foldMemoryOperand callback only. 990 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 991 MIB.addImm(StackMaps::IndirectMemRefOp); 992 MIB.addImm(MFI.getObjectSize(FI)); 993 MIB.add(MI->getOperand(OperIdx)); 994 MIB.addImm(0); 995 } else { 996 // direct-mem-ref tag, #FI, offset. 997 // Used by patchpoint, and direct alloca arguments to statepoints 998 MIB.addImm(StackMaps::DirectMemRefOp); 999 MIB.add(MI->getOperand(OperIdx)); 1000 MIB.addImm(0); 1001 } 1002 // Copy the operands after the frame index. 1003 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 1004 MIB.add(MI->getOperand(i)); 1005 1006 // Inherit previous memory operands. 1007 MIB.cloneMemRefs(*MI); 1008 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 1009 1010 // Add a new memory operand for this FI. 1011 assert(MFI.getObjectOffset(FI) != -1); 1012 1013 // Note: STATEPOINT MMOs are added during SelectionDAG. STACKMAP, and 1014 // PATCHPOINT should be updated to do the same. (TODO) 1015 if (MI->getOpcode() != TargetOpcode::STATEPOINT) { 1016 auto Flags = MachineMemOperand::MOLoad; 1017 MachineMemOperand *MMO = MF.getMachineMemOperand( 1018 MachinePointerInfo::getFixedStack(MF, FI), Flags, 1019 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 1020 MIB->addMemOperand(MF, MMO); 1021 } 1022 1023 // Replace the instruction and update the operand index. 1024 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1025 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1026 MI->eraseFromParent(); 1027 MI = MIB; 1028 } 1029 return MBB; 1030 } 1031 1032 MachineBasicBlock * 1033 TargetLoweringBase::emitXRayCustomEvent(MachineInstr &MI, 1034 MachineBasicBlock *MBB) const { 1035 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_EVENT_CALL && 1036 "Called emitXRayCustomEvent on the wrong MI!"); 1037 auto &MF = *MI.getMF(); 1038 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1039 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1040 MIB.add(MI.getOperand(OpIdx)); 1041 1042 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1043 MI.eraseFromParent(); 1044 return MBB; 1045 } 1046 1047 MachineBasicBlock * 1048 TargetLoweringBase::emitXRayTypedEvent(MachineInstr &MI, 1049 MachineBasicBlock *MBB) const { 1050 assert(MI.getOpcode() == TargetOpcode::PATCHABLE_TYPED_EVENT_CALL && 1051 "Called emitXRayTypedEvent on the wrong MI!"); 1052 auto &MF = *MI.getMF(); 1053 auto MIB = BuildMI(MF, MI.getDebugLoc(), MI.getDesc()); 1054 for (unsigned OpIdx = 0; OpIdx != MI.getNumOperands(); ++OpIdx) 1055 MIB.add(MI.getOperand(OpIdx)); 1056 1057 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1058 MI.eraseFromParent(); 1059 return MBB; 1060 } 1061 1062 /// findRepresentativeClass - Return the largest legal super-reg register class 1063 /// of the register class for the specified type and its associated "cost". 1064 // This function is in TargetLowering because it uses RegClassForVT which would 1065 // need to be moved to TargetRegisterInfo and would necessitate moving 1066 // isTypeLegal over as well - a massive change that would just require 1067 // TargetLowering having a TargetRegisterInfo class member that it would use. 1068 std::pair<const TargetRegisterClass *, uint8_t> 1069 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1070 MVT VT) const { 1071 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1072 if (!RC) 1073 return std::make_pair(RC, 0); 1074 1075 // Compute the set of all super-register classes. 1076 BitVector SuperRegRC(TRI->getNumRegClasses()); 1077 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1078 SuperRegRC.setBitsInMask(RCI.getMask()); 1079 1080 // Find the first legal register class with the largest spill size. 1081 const TargetRegisterClass *BestRC = RC; 1082 for (unsigned i : SuperRegRC.set_bits()) { 1083 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1084 // We want the largest possible spill size. 1085 if (TRI->getSpillSize(*SuperRC) <= TRI->getSpillSize(*BestRC)) 1086 continue; 1087 if (!isLegalRC(*TRI, *SuperRC)) 1088 continue; 1089 BestRC = SuperRC; 1090 } 1091 return std::make_pair(BestRC, 1); 1092 } 1093 1094 /// computeRegisterProperties - Once all of the register classes are added, 1095 /// this allows us to compute derived properties we expose. 1096 void TargetLoweringBase::computeRegisterProperties( 1097 const TargetRegisterInfo *TRI) { 1098 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1099 "Too many value types for ValueTypeActions to hold!"); 1100 1101 // Everything defaults to needing one register. 1102 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1103 NumRegistersForVT[i] = 1; 1104 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1105 } 1106 // ...except isVoid, which doesn't need any registers. 1107 NumRegistersForVT[MVT::isVoid] = 0; 1108 1109 // Find the largest integer register class. 1110 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1111 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1112 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1113 1114 // Every integer value type larger than this largest register takes twice as 1115 // many registers to represent as the previous ValueType. 1116 for (unsigned ExpandedReg = LargestIntReg + 1; 1117 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1118 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1119 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1120 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1121 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1122 TypeExpandInteger); 1123 } 1124 1125 // Inspect all of the ValueType's smaller than the largest integer 1126 // register to see which ones need promotion. 1127 unsigned LegalIntReg = LargestIntReg; 1128 for (unsigned IntReg = LargestIntReg - 1; 1129 IntReg >= (unsigned)MVT::i1; --IntReg) { 1130 MVT IVT = (MVT::SimpleValueType)IntReg; 1131 if (isTypeLegal(IVT)) { 1132 LegalIntReg = IntReg; 1133 } else { 1134 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1135 (MVT::SimpleValueType)LegalIntReg; 1136 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1137 } 1138 } 1139 1140 // ppcf128 type is really two f64's. 1141 if (!isTypeLegal(MVT::ppcf128)) { 1142 if (isTypeLegal(MVT::f64)) { 1143 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1144 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1145 TransformToType[MVT::ppcf128] = MVT::f64; 1146 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1147 } else { 1148 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1149 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1150 TransformToType[MVT::ppcf128] = MVT::i128; 1151 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1152 } 1153 } 1154 1155 // Decide how to handle f128. If the target does not have native f128 support, 1156 // expand it to i128 and we will be generating soft float library calls. 1157 if (!isTypeLegal(MVT::f128)) { 1158 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1159 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1160 TransformToType[MVT::f128] = MVT::i128; 1161 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1162 } 1163 1164 // Decide how to handle f64. If the target does not have native f64 support, 1165 // expand it to i64 and we will be generating soft float library calls. 1166 if (!isTypeLegal(MVT::f64)) { 1167 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1168 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1169 TransformToType[MVT::f64] = MVT::i64; 1170 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1171 } 1172 1173 // Decide how to handle f32. If the target does not have native f32 support, 1174 // expand it to i32 and we will be generating soft float library calls. 1175 if (!isTypeLegal(MVT::f32)) { 1176 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1177 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1178 TransformToType[MVT::f32] = MVT::i32; 1179 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1180 } 1181 1182 // Decide how to handle f16. If the target does not have native f16 support, 1183 // promote it to f32, because there are no f16 library calls (except for 1184 // conversions). 1185 if (!isTypeLegal(MVT::f16)) { 1186 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1187 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1188 TransformToType[MVT::f16] = MVT::f32; 1189 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1190 } 1191 1192 // Loop over all of the vector value types to see which need transformations. 1193 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1194 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1195 MVT VT = (MVT::SimpleValueType) i; 1196 if (isTypeLegal(VT)) 1197 continue; 1198 1199 MVT EltVT = VT.getVectorElementType(); 1200 unsigned NElts = VT.getVectorNumElements(); 1201 bool IsLegalWiderType = false; 1202 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1203 switch (PreferredAction) { 1204 case TypePromoteInteger: 1205 // Try to promote the elements of integer vectors. If no legal 1206 // promotion was found, fall through to the widen-vector method. 1207 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1208 MVT SVT = (MVT::SimpleValueType) nVT; 1209 // Promote vectors of integers to vectors with the same number 1210 // of elements, with a wider element type. 1211 if (SVT.getScalarSizeInBits() > EltVT.getSizeInBits() && 1212 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1213 TransformToType[i] = SVT; 1214 RegisterTypeForVT[i] = SVT; 1215 NumRegistersForVT[i] = 1; 1216 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1217 IsLegalWiderType = true; 1218 break; 1219 } 1220 } 1221 if (IsLegalWiderType) 1222 break; 1223 LLVM_FALLTHROUGH; 1224 1225 case TypeWidenVector: 1226 // Try to widen the vector. 1227 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1228 MVT SVT = (MVT::SimpleValueType) nVT; 1229 if (SVT.getVectorElementType() == EltVT 1230 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1231 TransformToType[i] = SVT; 1232 RegisterTypeForVT[i] = SVT; 1233 NumRegistersForVT[i] = 1; 1234 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1235 IsLegalWiderType = true; 1236 break; 1237 } 1238 } 1239 if (IsLegalWiderType) 1240 break; 1241 LLVM_FALLTHROUGH; 1242 1243 case TypeSplitVector: 1244 case TypeScalarizeVector: { 1245 MVT IntermediateVT; 1246 MVT RegisterVT; 1247 unsigned NumIntermediates; 1248 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1249 NumIntermediates, RegisterVT, this); 1250 RegisterTypeForVT[i] = RegisterVT; 1251 1252 MVT NVT = VT.getPow2VectorType(); 1253 if (NVT == VT) { 1254 // Type is already a power of 2. The default action is to split. 1255 TransformToType[i] = MVT::Other; 1256 if (PreferredAction == TypeScalarizeVector) 1257 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1258 else if (PreferredAction == TypeSplitVector) 1259 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1260 else 1261 // Set type action according to the number of elements. 1262 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1263 : TypeSplitVector); 1264 } else { 1265 TransformToType[i] = NVT; 1266 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1267 } 1268 break; 1269 } 1270 default: 1271 llvm_unreachable("Unknown vector legalization action!"); 1272 } 1273 } 1274 1275 // Determine the 'representative' register class for each value type. 1276 // An representative register class is the largest (meaning one which is 1277 // not a sub-register class / subreg register class) legal register class for 1278 // a group of value types. For example, on i386, i8, i16, and i32 1279 // representative would be GR32; while on x86_64 it's GR64. 1280 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1281 const TargetRegisterClass* RRC; 1282 uint8_t Cost; 1283 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1284 RepRegClassForVT[i] = RRC; 1285 RepRegClassCostForVT[i] = Cost; 1286 } 1287 } 1288 1289 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1290 EVT VT) const { 1291 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1292 return getPointerTy(DL).SimpleTy; 1293 } 1294 1295 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1296 return MVT::i32; // return the default value 1297 } 1298 1299 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1300 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1301 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1302 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1303 /// 1304 /// This method returns the number of registers needed, and the VT for each 1305 /// register. It also returns the VT and quantity of the intermediate values 1306 /// before they are promoted/expanded. 1307 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1308 EVT &IntermediateVT, 1309 unsigned &NumIntermediates, 1310 MVT &RegisterVT) const { 1311 unsigned NumElts = VT.getVectorNumElements(); 1312 1313 // If there is a wider vector type with the same element type as this one, 1314 // or a promoted vector type that has the same number of elements which 1315 // are wider, then we should convert to that legal vector type. 1316 // This handles things like <2 x float> -> <4 x float> and 1317 // <4 x i1> -> <4 x i32>. 1318 LegalizeTypeAction TA = getTypeAction(Context, VT); 1319 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1320 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1321 if (isTypeLegal(RegisterEVT)) { 1322 IntermediateVT = RegisterEVT; 1323 RegisterVT = RegisterEVT.getSimpleVT(); 1324 NumIntermediates = 1; 1325 return 1; 1326 } 1327 } 1328 1329 // Figure out the right, legal destination reg to copy into. 1330 EVT EltTy = VT.getVectorElementType(); 1331 1332 unsigned NumVectorRegs = 1; 1333 1334 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1335 // could break down into LHS/RHS like LegalizeDAG does. 1336 if (!isPowerOf2_32(NumElts)) { 1337 NumVectorRegs = NumElts; 1338 NumElts = 1; 1339 } 1340 1341 // Divide the input until we get to a supported size. This will always 1342 // end with a scalar if the target doesn't support vectors. 1343 while (NumElts > 1 && !isTypeLegal( 1344 EVT::getVectorVT(Context, EltTy, NumElts))) { 1345 NumElts >>= 1; 1346 NumVectorRegs <<= 1; 1347 } 1348 1349 NumIntermediates = NumVectorRegs; 1350 1351 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1352 if (!isTypeLegal(NewVT)) 1353 NewVT = EltTy; 1354 IntermediateVT = NewVT; 1355 1356 MVT DestVT = getRegisterType(Context, NewVT); 1357 RegisterVT = DestVT; 1358 unsigned NewVTSize = NewVT.getSizeInBits(); 1359 1360 // Convert sizes such as i33 to i64. 1361 if (!isPowerOf2_32(NewVTSize)) 1362 NewVTSize = NextPowerOf2(NewVTSize); 1363 1364 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1365 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1366 1367 // Otherwise, promotion or legal types use the same number of registers as 1368 // the vector decimated to the appropriate level. 1369 return NumVectorRegs; 1370 } 1371 1372 /// Get the EVTs and ArgFlags collections that represent the legalized return 1373 /// type of the given function. This does not require a DAG or a return value, 1374 /// and is suitable for use before any DAGs for the function are constructed. 1375 /// TODO: Move this out of TargetLowering.cpp. 1376 void llvm::GetReturnInfo(CallingConv::ID CC, Type *ReturnType, 1377 AttributeList attr, 1378 SmallVectorImpl<ISD::OutputArg> &Outs, 1379 const TargetLowering &TLI, const DataLayout &DL) { 1380 SmallVector<EVT, 4> ValueVTs; 1381 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1382 unsigned NumValues = ValueVTs.size(); 1383 if (NumValues == 0) return; 1384 1385 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1386 EVT VT = ValueVTs[j]; 1387 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1388 1389 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1390 ExtendKind = ISD::SIGN_EXTEND; 1391 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1392 ExtendKind = ISD::ZERO_EXTEND; 1393 1394 // FIXME: C calling convention requires the return type to be promoted to 1395 // at least 32-bit. But this is not necessary for non-C calling 1396 // conventions. The frontend should mark functions whose return values 1397 // require promoting with signext or zeroext attributes. 1398 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1399 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1400 if (VT.bitsLT(MinVT)) 1401 VT = MinVT; 1402 } 1403 1404 unsigned NumParts = 1405 TLI.getNumRegistersForCallingConv(ReturnType->getContext(), CC, VT); 1406 MVT PartVT = 1407 TLI.getRegisterTypeForCallingConv(ReturnType->getContext(), CC, VT); 1408 1409 // 'inreg' on function refers to return value 1410 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1411 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::InReg)) 1412 Flags.setInReg(); 1413 1414 // Propagate extension type if any 1415 if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) 1416 Flags.setSExt(); 1417 else if (attr.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt)) 1418 Flags.setZExt(); 1419 1420 for (unsigned i = 0; i < NumParts; ++i) 1421 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0)); 1422 } 1423 } 1424 1425 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1426 /// function arguments in the caller parameter area. This is the actual 1427 /// alignment, not its logarithm. 1428 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1429 const DataLayout &DL) const { 1430 return DL.getABITypeAlignment(Ty); 1431 } 1432 1433 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1434 const DataLayout &DL, EVT VT, 1435 unsigned AddrSpace, 1436 unsigned Alignment, 1437 bool *Fast) const { 1438 // Check if the specified alignment is sufficient based on the data layout. 1439 // TODO: While using the data layout works in practice, a better solution 1440 // would be to implement this check directly (make this a virtual function). 1441 // For example, the ABI alignment may change based on software platform while 1442 // this function should only be affected by hardware implementation. 1443 Type *Ty = VT.getTypeForEVT(Context); 1444 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1445 // Assume that an access that meets the ABI-specified alignment is fast. 1446 if (Fast != nullptr) 1447 *Fast = true; 1448 return true; 1449 } 1450 1451 // This is a misaligned access. 1452 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast); 1453 } 1454 1455 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1456 return BranchProbability(MinPercentageForPredictableBranch, 100); 1457 } 1458 1459 //===----------------------------------------------------------------------===// 1460 // TargetTransformInfo Helpers 1461 //===----------------------------------------------------------------------===// 1462 1463 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1464 enum InstructionOpcodes { 1465 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1466 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1467 #include "llvm/IR/Instruction.def" 1468 }; 1469 switch (static_cast<InstructionOpcodes>(Opcode)) { 1470 case Ret: return 0; 1471 case Br: return 0; 1472 case Switch: return 0; 1473 case IndirectBr: return 0; 1474 case Invoke: return 0; 1475 case CallBr: return 0; 1476 case Resume: return 0; 1477 case Unreachable: return 0; 1478 case CleanupRet: return 0; 1479 case CatchRet: return 0; 1480 case CatchPad: return 0; 1481 case CatchSwitch: return 0; 1482 case CleanupPad: return 0; 1483 case FNeg: return ISD::FNEG; 1484 case Add: return ISD::ADD; 1485 case FAdd: return ISD::FADD; 1486 case Sub: return ISD::SUB; 1487 case FSub: return ISD::FSUB; 1488 case Mul: return ISD::MUL; 1489 case FMul: return ISD::FMUL; 1490 case UDiv: return ISD::UDIV; 1491 case SDiv: return ISD::SDIV; 1492 case FDiv: return ISD::FDIV; 1493 case URem: return ISD::UREM; 1494 case SRem: return ISD::SREM; 1495 case FRem: return ISD::FREM; 1496 case Shl: return ISD::SHL; 1497 case LShr: return ISD::SRL; 1498 case AShr: return ISD::SRA; 1499 case And: return ISD::AND; 1500 case Or: return ISD::OR; 1501 case Xor: return ISD::XOR; 1502 case Alloca: return 0; 1503 case Load: return ISD::LOAD; 1504 case Store: return ISD::STORE; 1505 case GetElementPtr: return 0; 1506 case Fence: return 0; 1507 case AtomicCmpXchg: return 0; 1508 case AtomicRMW: return 0; 1509 case Trunc: return ISD::TRUNCATE; 1510 case ZExt: return ISD::ZERO_EXTEND; 1511 case SExt: return ISD::SIGN_EXTEND; 1512 case FPToUI: return ISD::FP_TO_UINT; 1513 case FPToSI: return ISD::FP_TO_SINT; 1514 case UIToFP: return ISD::UINT_TO_FP; 1515 case SIToFP: return ISD::SINT_TO_FP; 1516 case FPTrunc: return ISD::FP_ROUND; 1517 case FPExt: return ISD::FP_EXTEND; 1518 case PtrToInt: return ISD::BITCAST; 1519 case IntToPtr: return ISD::BITCAST; 1520 case BitCast: return ISD::BITCAST; 1521 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1522 case ICmp: return ISD::SETCC; 1523 case FCmp: return ISD::SETCC; 1524 case PHI: return 0; 1525 case Call: return 0; 1526 case Select: return ISD::SELECT; 1527 case UserOp1: return 0; 1528 case UserOp2: return 0; 1529 case VAArg: return 0; 1530 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1531 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1532 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1533 case ExtractValue: return ISD::MERGE_VALUES; 1534 case InsertValue: return ISD::MERGE_VALUES; 1535 case LandingPad: return 0; 1536 } 1537 1538 llvm_unreachable("Unknown instruction type encountered!"); 1539 } 1540 1541 std::pair<int, MVT> 1542 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1543 Type *Ty) const { 1544 LLVMContext &C = Ty->getContext(); 1545 EVT MTy = getValueType(DL, Ty); 1546 1547 int Cost = 1; 1548 // We keep legalizing the type until we find a legal kind. We assume that 1549 // the only operation that costs anything is the split. After splitting 1550 // we need to handle two types. 1551 while (true) { 1552 LegalizeKind LK = getTypeConversion(C, MTy); 1553 1554 if (LK.first == TypeLegal) 1555 return std::make_pair(Cost, MTy.getSimpleVT()); 1556 1557 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1558 Cost *= 2; 1559 1560 // Do not loop with f128 type. 1561 if (MTy == LK.second) 1562 return std::make_pair(Cost, MTy.getSimpleVT()); 1563 1564 // Keep legalizing the type. 1565 MTy = LK.second; 1566 } 1567 } 1568 1569 Value *TargetLoweringBase::getDefaultSafeStackPointerLocation(IRBuilder<> &IRB, 1570 bool UseTLS) const { 1571 // compiler-rt provides a variable with a magic name. Targets that do not 1572 // link with compiler-rt may also provide such a variable. 1573 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1574 const char *UnsafeStackPtrVar = "__safestack_unsafe_stack_ptr"; 1575 auto UnsafeStackPtr = 1576 dyn_cast_or_null<GlobalVariable>(M->getNamedValue(UnsafeStackPtrVar)); 1577 1578 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1579 1580 if (!UnsafeStackPtr) { 1581 auto TLSModel = UseTLS ? 1582 GlobalValue::InitialExecTLSModel : 1583 GlobalValue::NotThreadLocal; 1584 // The global variable is not defined yet, define it ourselves. 1585 // We use the initial-exec TLS model because we do not support the 1586 // variable living anywhere other than in the main executable. 1587 UnsafeStackPtr = new GlobalVariable( 1588 *M, StackPtrTy, false, GlobalValue::ExternalLinkage, nullptr, 1589 UnsafeStackPtrVar, nullptr, TLSModel); 1590 } else { 1591 // The variable exists, check its type and attributes. 1592 if (UnsafeStackPtr->getValueType() != StackPtrTy) 1593 report_fatal_error(Twine(UnsafeStackPtrVar) + " must have void* type"); 1594 if (UseTLS != UnsafeStackPtr->isThreadLocal()) 1595 report_fatal_error(Twine(UnsafeStackPtrVar) + " must " + 1596 (UseTLS ? "" : "not ") + "be thread-local"); 1597 } 1598 return UnsafeStackPtr; 1599 } 1600 1601 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1602 if (!TM.getTargetTriple().isAndroid()) 1603 return getDefaultSafeStackPointerLocation(IRB, true); 1604 1605 // Android provides a libc function to retrieve the address of the current 1606 // thread's unsafe stack pointer. 1607 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1608 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1609 FunctionCallee Fn = M->getOrInsertFunction("__safestack_pointer_address", 1610 StackPtrTy->getPointerTo(0)); 1611 return IRB.CreateCall(Fn); 1612 } 1613 1614 //===----------------------------------------------------------------------===// 1615 // Loop Strength Reduction hooks 1616 //===----------------------------------------------------------------------===// 1617 1618 /// isLegalAddressingMode - Return true if the addressing mode represented 1619 /// by AM is legal for this target, for a load/store of the specified type. 1620 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1621 const AddrMode &AM, Type *Ty, 1622 unsigned AS, Instruction *I) const { 1623 // The default implementation of this implements a conservative RISCy, r+r and 1624 // r+i addr mode. 1625 1626 // Allows a sign-extended 16-bit immediate field. 1627 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1628 return false; 1629 1630 // No global is ever allowed as a base. 1631 if (AM.BaseGV) 1632 return false; 1633 1634 // Only support r+r, 1635 switch (AM.Scale) { 1636 case 0: // "r+i" or just "i", depending on HasBaseReg. 1637 break; 1638 case 1: 1639 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1640 return false; 1641 // Otherwise we have r+r or r+i. 1642 break; 1643 case 2: 1644 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1645 return false; 1646 // Allow 2*r as r+r. 1647 break; 1648 default: // Don't allow n * r 1649 return false; 1650 } 1651 1652 return true; 1653 } 1654 1655 //===----------------------------------------------------------------------===// 1656 // Stack Protector 1657 //===----------------------------------------------------------------------===// 1658 1659 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1660 // so that SelectionDAG handle SSP. 1661 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1662 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1663 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1664 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1665 return M.getOrInsertGlobal("__guard_local", PtrTy); 1666 } 1667 return nullptr; 1668 } 1669 1670 // Currently only support "standard" __stack_chk_guard. 1671 // TODO: add LOAD_STACK_GUARD support. 1672 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1673 if (!M.getNamedValue("__stack_chk_guard")) 1674 new GlobalVariable(M, Type::getInt8PtrTy(M.getContext()), false, 1675 GlobalVariable::ExternalLinkage, 1676 nullptr, "__stack_chk_guard"); 1677 } 1678 1679 // Currently only support "standard" __stack_chk_guard. 1680 // TODO: add LOAD_STACK_GUARD support. 1681 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1682 return M.getNamedValue("__stack_chk_guard"); 1683 } 1684 1685 Function *TargetLoweringBase::getSSPStackGuardCheck(const Module &M) const { 1686 return nullptr; 1687 } 1688 1689 unsigned TargetLoweringBase::getMinimumJumpTableEntries() const { 1690 return MinimumJumpTableEntries; 1691 } 1692 1693 void TargetLoweringBase::setMinimumJumpTableEntries(unsigned Val) { 1694 MinimumJumpTableEntries = Val; 1695 } 1696 1697 unsigned TargetLoweringBase::getMinimumJumpTableDensity(bool OptForSize) const { 1698 return OptForSize ? OptsizeJumpTableDensity : JumpTableDensity; 1699 } 1700 1701 unsigned TargetLoweringBase::getMaximumJumpTableSize() const { 1702 return MaximumJumpTableSize; 1703 } 1704 1705 void TargetLoweringBase::setMaximumJumpTableSize(unsigned Val) { 1706 MaximumJumpTableSize = Val; 1707 } 1708 1709 //===----------------------------------------------------------------------===// 1710 // Reciprocal Estimates 1711 //===----------------------------------------------------------------------===// 1712 1713 /// Get the reciprocal estimate attribute string for a function that will 1714 /// override the target defaults. 1715 static StringRef getRecipEstimateForFunc(MachineFunction &MF) { 1716 const Function &F = MF.getFunction(); 1717 return F.getFnAttribute("reciprocal-estimates").getValueAsString(); 1718 } 1719 1720 /// Construct a string for the given reciprocal operation of the given type. 1721 /// This string should match the corresponding option to the front-end's 1722 /// "-mrecip" flag assuming those strings have been passed through in an 1723 /// attribute string. For example, "vec-divf" for a division of a vXf32. 1724 static std::string getReciprocalOpName(bool IsSqrt, EVT VT) { 1725 std::string Name = VT.isVector() ? "vec-" : ""; 1726 1727 Name += IsSqrt ? "sqrt" : "div"; 1728 1729 // TODO: Handle "half" or other float types? 1730 if (VT.getScalarType() == MVT::f64) { 1731 Name += "d"; 1732 } else { 1733 assert(VT.getScalarType() == MVT::f32 && 1734 "Unexpected FP type for reciprocal estimate"); 1735 Name += "f"; 1736 } 1737 1738 return Name; 1739 } 1740 1741 /// Return the character position and value (a single numeric character) of a 1742 /// customized refinement operation in the input string if it exists. Return 1743 /// false if there is no customized refinement step count. 1744 static bool parseRefinementStep(StringRef In, size_t &Position, 1745 uint8_t &Value) { 1746 const char RefStepToken = ':'; 1747 Position = In.find(RefStepToken); 1748 if (Position == StringRef::npos) 1749 return false; 1750 1751 StringRef RefStepString = In.substr(Position + 1); 1752 // Allow exactly one numeric character for the additional refinement 1753 // step parameter. 1754 if (RefStepString.size() == 1) { 1755 char RefStepChar = RefStepString[0]; 1756 if (RefStepChar >= '0' && RefStepChar <= '9') { 1757 Value = RefStepChar - '0'; 1758 return true; 1759 } 1760 } 1761 report_fatal_error("Invalid refinement step for -recip."); 1762 } 1763 1764 /// For the input attribute string, return one of the ReciprocalEstimate enum 1765 /// status values (enabled, disabled, or not specified) for this operation on 1766 /// the specified data type. 1767 static int getOpEnabled(bool IsSqrt, EVT VT, StringRef Override) { 1768 if (Override.empty()) 1769 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1770 1771 SmallVector<StringRef, 4> OverrideVector; 1772 Override.split(OverrideVector, ','); 1773 unsigned NumArgs = OverrideVector.size(); 1774 1775 // Check if "all", "none", or "default" was specified. 1776 if (NumArgs == 1) { 1777 // Look for an optional setting of the number of refinement steps needed 1778 // for this type of reciprocal operation. 1779 size_t RefPos; 1780 uint8_t RefSteps; 1781 if (parseRefinementStep(Override, RefPos, RefSteps)) { 1782 // Split the string for further processing. 1783 Override = Override.substr(0, RefPos); 1784 } 1785 1786 // All reciprocal types are enabled. 1787 if (Override == "all") 1788 return TargetLoweringBase::ReciprocalEstimate::Enabled; 1789 1790 // All reciprocal types are disabled. 1791 if (Override == "none") 1792 return TargetLoweringBase::ReciprocalEstimate::Disabled; 1793 1794 // Target defaults for enablement are used. 1795 if (Override == "default") 1796 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1797 } 1798 1799 // The attribute string may omit the size suffix ('f'/'d'). 1800 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1801 std::string VTNameNoSize = VTName; 1802 VTNameNoSize.pop_back(); 1803 static const char DisabledPrefix = '!'; 1804 1805 for (StringRef RecipType : OverrideVector) { 1806 size_t RefPos; 1807 uint8_t RefSteps; 1808 if (parseRefinementStep(RecipType, RefPos, RefSteps)) 1809 RecipType = RecipType.substr(0, RefPos); 1810 1811 // Ignore the disablement token for string matching. 1812 bool IsDisabled = RecipType[0] == DisabledPrefix; 1813 if (IsDisabled) 1814 RecipType = RecipType.substr(1); 1815 1816 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1817 return IsDisabled ? TargetLoweringBase::ReciprocalEstimate::Disabled 1818 : TargetLoweringBase::ReciprocalEstimate::Enabled; 1819 } 1820 1821 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1822 } 1823 1824 /// For the input attribute string, return the customized refinement step count 1825 /// for this operation on the specified data type. If the step count does not 1826 /// exist, return the ReciprocalEstimate enum value for unspecified. 1827 static int getOpRefinementSteps(bool IsSqrt, EVT VT, StringRef Override) { 1828 if (Override.empty()) 1829 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1830 1831 SmallVector<StringRef, 4> OverrideVector; 1832 Override.split(OverrideVector, ','); 1833 unsigned NumArgs = OverrideVector.size(); 1834 1835 // Check if "all", "default", or "none" was specified. 1836 if (NumArgs == 1) { 1837 // Look for an optional setting of the number of refinement steps needed 1838 // for this type of reciprocal operation. 1839 size_t RefPos; 1840 uint8_t RefSteps; 1841 if (!parseRefinementStep(Override, RefPos, RefSteps)) 1842 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1843 1844 // Split the string for further processing. 1845 Override = Override.substr(0, RefPos); 1846 assert(Override != "none" && 1847 "Disabled reciprocals, but specifed refinement steps?"); 1848 1849 // If this is a general override, return the specified number of steps. 1850 if (Override == "all" || Override == "default") 1851 return RefSteps; 1852 } 1853 1854 // The attribute string may omit the size suffix ('f'/'d'). 1855 std::string VTName = getReciprocalOpName(IsSqrt, VT); 1856 std::string VTNameNoSize = VTName; 1857 VTNameNoSize.pop_back(); 1858 1859 for (StringRef RecipType : OverrideVector) { 1860 size_t RefPos; 1861 uint8_t RefSteps; 1862 if (!parseRefinementStep(RecipType, RefPos, RefSteps)) 1863 continue; 1864 1865 RecipType = RecipType.substr(0, RefPos); 1866 if (RecipType.equals(VTName) || RecipType.equals(VTNameNoSize)) 1867 return RefSteps; 1868 } 1869 1870 return TargetLoweringBase::ReciprocalEstimate::Unspecified; 1871 } 1872 1873 int TargetLoweringBase::getRecipEstimateSqrtEnabled(EVT VT, 1874 MachineFunction &MF) const { 1875 return getOpEnabled(true, VT, getRecipEstimateForFunc(MF)); 1876 } 1877 1878 int TargetLoweringBase::getRecipEstimateDivEnabled(EVT VT, 1879 MachineFunction &MF) const { 1880 return getOpEnabled(false, VT, getRecipEstimateForFunc(MF)); 1881 } 1882 1883 int TargetLoweringBase::getSqrtRefinementSteps(EVT VT, 1884 MachineFunction &MF) const { 1885 return getOpRefinementSteps(true, VT, getRecipEstimateForFunc(MF)); 1886 } 1887 1888 int TargetLoweringBase::getDivRefinementSteps(EVT VT, 1889 MachineFunction &MF) const { 1890 return getOpRefinementSteps(false, VT, getRecipEstimateForFunc(MF)); 1891 } 1892 1893 void TargetLoweringBase::finalizeLowering(MachineFunction &MF) const { 1894 MF.getRegInfo().freezeReservedRegs(MF); 1895 } 1896